1. hobbies and interests
2. games
3. video and computer games

Psychological Bulletin
2010, Vol. 136, No. 2, 151–173
© 2010 American Psychological Association
0033-2909/10/$12.00 DOI: 10.1037/a0018251
Violent Video Game Effects on Aggression, Empathy, and Prosocial
Behavior in Eastern and Western Countries: A Meta-Analytic Review
Craig A. Anderson
Akiko Shibuya
Iowa State University
Keio University
Nobuko Ihori
Edward L. Swing
Ochanomizu University
Iowa State University
Brad J. Bushman
Akira Sakamoto
VU University and University of Michigan
Ochanomizu University
Hannah R. Rothstein
Muniba Saleem
City University of New York
Iowa State University
Meta-analytic procedures were used to test the effects of violent video games on aggressive behavior,
aggressive cognition, aggressive affect, physiological arousal, empathy/desensitization, and prosocial behavior. Unique features of this meta-analytic review include (a) more restrictive methodological quality inclusion
criteria than in past meta-analyses; (b) cross-cultural comparisons; (c) longitudinal studies for all outcomes
except physiological arousal; (d) conservative statistical controls; (e) multiple moderator analyses; and (f)
sensitivity analyses. Social– cognitive models and cultural differences between Japan and Western countries
were used to generate theory-based predictions. Meta-analyses yielded significant effects for all 6 outcome
variables. The pattern of results for different outcomes and research designs (experimental, cross-sectional,
longitudinal) fit theoretical predictions well. The evidence strongly suggests that exposure to violent video
games is a causal risk factor for increased aggressive behavior, aggressive cognition, and aggressive affect and
for decreased empathy and prosocial behavior. Moderator analyses revealed significant research design
effects, weak evidence of cultural differences in susceptibility and type of measurement effects, and no
evidence of sex differences in susceptibility. Results of various sensitivity analyses revealed these effects to
be robust, with little evidence of selection (publication) bias.
Keywords: media violence, aggression, video games, empathy and desensitization, prosocial behavior
Supplemental materials: http://dx.doi.org/10.1037/a0018251.supp
People of all ages in most modern countries get a heavy dose of
violent media, especially in TV programs, films, and video games
(e.g., Comstock & Scharrer, 2007; Gentile, 2003; Gentile, Saleem,
& Anderson, 2007; Kirsh, 2006; Singer & Singer, 2001). Potential
harmful effects of media violence have been scrutinized for over
six decades, and considerable consensus has been reached on
several of the most important issues. As stated by a recent panel of
experts assembled by the U.S. Surgeon General, “Research on
violent television and films, video games, and music reveals unequivocal evidence that media violence increases the likelihood of
aggressive and violent behavior in both immediate and long-term
contexts” (Anderson et al., 2003, p. 81). Numerous reports by
professional health associations (e.g., American Academy of Pediatrics, American Psychological Association, Australian College
of Paediatrics, Canadian Paediatric Society) and government
health agencies (e.g., U.S. Office of the Surgeon General, U.S.
Department of Health and Human Services) have reached the same
conclusion after reviewing the available scientific evidence (Gentile et al., 2007; Ontario Office for Victims of Crime, 2004).
You know what’s really exciting about video games is you don’t just
interact with the game physically—you’re not just moving your hand
on a joystick, but you’re asked to interact with the game psychologically and emotionally as well. You’re not just watching the characters
on screen; you’re becoming those characters.
—Nina Huntemann, Game Over
Craig A. Anderson, Edward L. Swing, and Muniba Saleem, Center for
the Study of Violence, Department of Psychology, Iowa State University;
Akiko Shibuya, Institute for Media and Communications Research, Keio
University, Tokyo, Japan; Nobuko Ihori and Akira Sakamoto, Department
of Psychology, Ochanomizu University, Tokyo, Japan; Brad J. Bushman,
Vrije Universiteit, Amsterdam, the Netherlands, and Institute for Social
Research, University of Michigan; Hannah R. Rothstein, Department of
Management, Baruch College, City University of New York.
Correspondence concerning this article should be addressed to Craig A.
Anderson, Department of Psychology, W112 Lagomarcino Hall, Iowa
State University, Ames, IA 50011. E-mail: caa@iastate.edu
151
ANDERSON ET AL.
152
The majority of media violence studies have focused on violent
television and film effects, and most have been conducted in
Western countries, especially the United States. There are theoretical reasons to expect that type of media (e.g., newspapers, literature, comic books, graphic novels, television, film, video games,
music) and culture will moderate violent media effects. For example, watching the Lord of the Rings films should increase aggressive tendencies more than reading the books because of the higher
concentration and glorification of violence in the films.
Similarly, cultural factors may either exacerbate or reduce violent media effects for both statistical and psychological reasons.
For example, the context of violence on Japanese television is very
different from that on U.S. television, even though the total amount
of violence shown is similar (Kodaira, 1998). Japanese TV tends to
portray violent actions and their consequences much more vividly,
with a particular emphasis on the suffering of the victims. This
might explain why the effects of TV violence on aggression
sometimes appear smaller in Japan than in the United States.
Other multinational research has found considerable variation in
access to and content of “violent television” and a few differences
in observed effects (Huesmann & Eron, 1986; Huesmann, Lagerspetz, & Eron, 1984). For example, within Israel there were
significant correlations between TV violence viewing and children’s aggression for urban children but not for rural children
being raised on a kibbutz, where socialization is conducted in a
communal manner (Bachrach, 1986). What is currently unclear is
the extent to which the occasional cross-cultural differences in
media violence effects result from cross-cultural differences in the
content of their violent media (a type of artifact), true differences
in media violence effects (perhaps communal, collectivist, or politically unstable countries are less susceptible), or a combination
of the two.
Video Game Violence
Past Findings
Video game violence is the new kid on the media violence
block, having emerged in the late 1980s and early 1990s. Currently, one can play video games on computers, consoles (e.g.,
Xbox 360, PlayStation, Wii), handhelds (e.g., Nintendo DS), computers, iPods, personal digital assistants, and mobile telephones.
Because video game technology is relatively new, there are fewer
empirical studies on video game violence than on TV and film
violence. Nonetheless, several meta-analytic reviews have reported significant harmful effects of exposure to violent video
games, both in short-term experimental studies and in crosssectional correlational studies (Anderson, 2004; Anderson &
Bushman, 2001; Anderson et al., 2004; Sherry, 2001). Briefly,
these reviews found that across these two research designs, exposure to violent video games is associated with higher levels of
aggressive behavior, aggressive cognition, aggressive affect, and
physiological arousal and with lower levels of prosocial behavior.
The earliest meta-analyses reported average effects on aggressive
behavior of r⫹ ⫽ .15 (K ⫽ 25, N ⫽ 2,722; Sherry, 2001) and
r⫹ ⫽ .19 (K ⫽ 33, N ⫽ 3,033; Anderson & Bushman, 2001).
Anderson (2004) found an average effect size of r⫹ ⫽ .20 (K ⫽
32, N ⫽ 5,240) when all relevant studies were included and a
larger effect when more stringent methodological criteria were
applied, r⫹ ⫽ .26 (K ⫽ 17, N ⫽ 2,731).1 In general, the violent
video game research mirrors findings from the violent TV and film
research, with some evidence that the violent video game effects
may be somewhat larger (Anderson, Gentile, & Buckley, 2007;
Polman, Orobio de Castro, & Van Aken, 2008).
Recent Skepticism
However, three recent meta-analyses by the same author, each
using a very small set of available studies, have suggested that the
effects of violent video games on aggression have been substantially overestimated because of publication bias (Ferguson, 2007a,
2007b; Ferguson & Kilburn, 2009) and that therefore there is
little-to-no evidence of a violent video game effect on aggression.
However, these three meta-analyses have numerous problems that
call into question their results and conclusions. For example,
counter to widely accepted procedures for reducing the impact of
publication bias, only published articles were included in the
analyses and then procedures for addressing publication bias were
misinterpreted. Also, studies published prior to 1995 were ignored
and a large number of studies published since that time apparently
were missed.
The text on publication bias cited by Ferguson (2007a; Rothstein, Sutton, & Borenstein, 2005) specifically recommends that
the primary way to assure that meta-analytic results will not be
affected by publication bias is to conduct a search for relevant
studies that is thorough, systematic, unbiased, transparent, and
clearly documented. Authors are told to include book chapters,
dissertations, conference papers, and unpublished manuscripts that
meet the inclusion criteria for the meta-analysis, because this is
widely viewed as the best way to ameliorate publication bias.
Ferguson (2007a, 2007b; Ferguson & Kilburn, 2009) used the
trim and fill method to estimate the “true” effect size corrected for
publication bias. The originators of the trim and fill method
(Duval, 2005; Duval & Tweedie, 2000a) have cautioned that the
“adjusted” estimate of an effect using imputed studies provided by
trim and fill should not be taken as the “true” effect, because it is
based on imputed data points (that do not really exist). Trim and
fill provides a useful sensitivity analysis that assesses the potential
impact of missing studies on the results of a meta-analysis by
examining the degree of divergence between the original effectsize estimate and the trim and fill adjusted effect-size estimate.
It has also been widely cautioned that because trim and fill and
some other techniques for assessing publication bias are based on
an association between effect size and sample size, other explanations of this association should be considered. For example, effect
sizes in experimental studies may be larger than those in crosssectional or longitudinal studies due to the reduced error variance
that results from tight experimental controls; researchers may
know this and therefore may intentionally plan to use larger
sample sizes when conducting nonexperimental studies. Similarly,
in some research contexts with very large sample sizes (e.g.,
national surveys) a researcher may have to use less precise measures (e.g., fewer items) that result in smaller effect sizes. In sum,
1
The Anderson et al. (2004) analysis differed only slightly from Anderson (2004) and yielded an almost identical effect for the methodologically
better studies, r⫹ ⫽ .27 (K ⫽ 18, N ⫽ 3,537).
VIOLENT VIDEO GAME EFFECTS
it is possible that the effects in the studies with small samples
really are larger than those in the studies with large samples (cf.
Sterne and Egger, 2005).
In addition, the meta-analyses published by Ferguson are not
independent of each other because they use highly overlapping
subsets of the same small sample of studies, which includes at least
one study that does not even have a valid measure of aggressive
behavior (i.e., Williams & Skoric, 2005). For example, the Ferguson (2007b) meta-analysis used data from 17 articles, 14 of which
were used in Ferguson (2007a), making the two meta-analyses
largely redundant.2 The average effect-size estimates computed by
Ferguson (rs ⫽ .29, .14, and .15 for Ferguson 2007a, 2007b, and
Ferguson & Kilburn, 2009, respectively) before “correcting” for
publication bias are very similar in magnitude to those computed
by other researchers.
Need for a New Meta-Analysis
Thus, there is some inconsistency between the recent metaanalyses conducted by Ferguson (2007a, 2007b; Ferguson & Kilburn, 2009) and most of the published research and earlier, more
comprehensive meta-analyses on media violence effects. Clearly,
all agree that prior meta-analyses have not answered all relevant
questions about violent video game effects. Furthermore, there has
been an explosion of research on violent video game effects since
the last comprehensive meta-analysis was published in 2004. For
example, none of the prior meta-analytic reviews of violent video
game effects included longitudinal studies because none existed,
but now several such studies are available. Past meta-analyses also
frequently included cross-sectional studies in which sex differences were not statistically controlled. Although there are both
theoretical and methodological reasons for not using partial correlations, it certainly is of interest to know whether the average
effect size is reliably different from zero when sex has been
controlled. A sufficient number of studies now exists to allow
meaningful tests of this question.
Other important questions could not be tested in prior metaanalyses because of the small number of available studies. For
example, does player perspective (first person vs. third person)
influence the magnitude of violent video game effects? Does
killing human targets yield larger effects than killing nonhuman
targets? Are younger game players more affected than older ones?
Furthermore, almost all of the studies reviewed in prior metaanalyses came from U.S. samples or from similar Western
individualistic-culture samples (e.g., Australia, Germany, the
Netherlands, United Kingdom). Thus, the possibility that video
game violence effects might be smaller (or larger) in collectivistic
societies than in individualistic societies has never been explored.
Indeed, it was the combination of availability of Japanese studies
(following a visit to Japan in 2003 by Craig A. Anderson), the
explosion of research in this domain, and the publication of several
longitudinal studies that inspired us to begin the present metaanalysis.
Cultural Differences in Aggression
Aggression rates differ greatly across countries and cultures;
cross-national comparisons have implicated various cultural variables as possible contributors to these differences. For example, an
153
analysis of peer-directed aggression in 28 countries found that “in
general, cultures characterized by collectivistic values, high moral
discipline, a high level of egalitarian commitment, low uncertainty
avoidance, and which emphasize values that are heavily Confucian
showed lower levels of aggression than their counterparts”
(Bergeron & Schneider, 2005, p. 116).3
However, the rank order of countries by aggression rates varies
from one measure of aggression to another. The United States has
a higher homicide rate than do many industrialized countries in
Europe and Asia but similar or lower rates of other forms of
violent crime, such as assault. For example, the average annual
homicide rate per hundred thousand for 1998 –2000 was almost
400% larger in the United States than in England and Wales (5.87
vs. 1.50; Barclay & Tavares, 2002). But in this time period, the rate
for all violent crimes was almost 250% greater in England and
Wales than in the United States (1,295 vs. 536; computed from
data reported in Barclay & Tavares, 2002).4
Japan is generally considered to be a relatively peaceful society.
It has lower rates of homicide (1.06) and violent crime (39) than
does the United States or most Western countries (Barclay &
Tavares, 2002). Japan is also a more collectivistic society, emphasizing high moral discipline, egalitarian values, and Confucian
values of peace and nonviolence.
One argument frequently offered by those who claim that media
violence doesn’t increase aggressive tendencies is that Japan has
high levels of media violence but low overall levels of violent
crime. If media violence is truly a causal risk factor for violence
and aggression, so the argument goes, Japan should have a high
violent crime rate. There are multiple problems with this argument,
of course. Perhaps the most obvious problem is that exposure to
violent media is not the only important risk factor (DeLisi, 2005).
Japan differs from the United States and other Western nations on
many known causal risk factors for aggression and violence, such
as easy access to firearms.
There are at least five reasons to expect smaller media violence
effect sizes in Japan (and similar Eastern societies) than in Western
societies. First, a relatively smaller effect size may result from
differences in how violence is contextualized in Japanese versus
U.S. media. Today, global boundaries do not exist when it comes
to video games. The most popular video games are played in many
countries, under different titles and with different languages.
Nonetheless, the contexts of violence in video games played most
frequently in Japan can be different from the contexts in video
games played most frequently in the United States. Whereas action
and sports games are the most popular genre in the United States
and Western countries, role-playing games are the most popular
2
We thank Christopher Ferguson for providing the list of articles used
in his three meta-analyses.
3
Note that of the violent video game studies we located from Eastern
cultures, the vast majority came from Japan. Indeed, there were only two
studies from other Eastern cultures, one from Singapore and the other from
China.
4
Differences in crime definitions and reporting method may account for
some portion of this reversal, but most scholars in the field agree that
violence rates in the United Kingdom are higher than in the United States,
with the exception of that for homicide, which is considerably higher in the
United States. A common explanation for the high homicide rate in the
United States is the easy availability of guns, especially handguns.
154
ANDERSON ET AL.
genre in Japan (Yahiro, 2005). Japanese role-playing games often
involve text reading, patience, and cooperative fights against
computer-controlled characters, and the contexts of the violent
video games that children and adolescents are exposed to in Japan
are not the same as those in the West. Second, people in Japan are
more likely to pay attention to situational contexts than are people
in Western countries (e.g., Masuda & Nisbett, 2001; Nisbett, Peng,
Choi, & Norenzayan, 2001). A third reason concerns cultural
differences in the meaning, experience, and processing of emotions and their emotion–action linkages. As noted by Mesquita and
Leu (2007), “Whereas people in independent contexts view emotional situations mainly from their own perspective . . . people in
interdependent contexts assess the emotional meaning from the
perspective of other people or a generalized other” (p. 739). One
result of this difference in perspective is that people from Japan
report being less likely to respond aggressively to an offense or
insult than do people in Western cultures. Other research similarly
suggests that Japanese culture tends to foster socially engaging
emotions, whereas Western culture tends to foster socially disengaging emotions (e.g., Kitayama, Mesquita, & Karasawa, 2006).
Similarly, research on ideal affect (what one typically would like
to feel) suggests that easterners are more likely to have adjustment
goals, whereas westerners are more likely to have influence goals
(Tsai, Knutson, & Fung, 2006). A fourth difference concerns the
context in which video games are played. One unpublished study
(Kodomo no taiken katsudo kenkyukai, 2000) found that considerably fewer Japanese and South Korean fifth graders had their
own TV sets (14% and 11%, respectively) than did American,
British, and German fifth graders (39%, 69%, and 29%, respectively). Similar results were obtained for eighth graders (28% and
10% vs. 63%, 68%, and 62%). This suggests that Japanese youths
may be more likely than Western youths to play their video games
in public space, where parents can watch and monitor what they
play. Research shows that parental involvement may reduce violent video game effects (e.g., Anderson et al., 2007). Research has
also found that the number of friends was not different for frequent
versus infrequent gamers in Japan, but that in the United Kingdom
frequent gamers had fewer friends than did infrequent gamers
(Colwell & Kato, 2003). Again, this suggests important context
differences between East and West that might moderate video
game effect sizes. Thus, violent video game effects on aggression
and related outcome variables may be smaller and more complex
in Japan than in Western countries.
On the other hand, most basic emotion and behavior processes
are universal. For example, Frijda, Markam, Sato, and Wiers
(1995) studied the action readiness after emotional experiences in
Dutch, Indonesian, and Japanese participants and found that five
factors were quite similar across cultures, including a factor labeled moving against. (There were some nonuniversal factors as
well.) Similarly, there are numerous cross-cultural differences in
average Big Five personality traits, some of which suggest that
Eastern collectivist cultures might be more susceptible to media
violence effects, others of which suggest the opposite (Schmitt,
Allik, McCrae, & Benet-Martinez, 2007). Thus, there also are
reasons to believe that media violence effects may well be fairly
similar across cultures or even larger in Japan and Eastern cultures.
In the present meta-analysis, we investigated the possibility that
effect sizes might differ between Western cultures (primarily the
United States) and Eastern cultures (primarily Japan).
Additional Theoretical Considerations
Over the last 45 years, an array of social– cognitive models of
aggression has systematically improved the field’s understanding
of the processes involved in the instigation of aggressive behavior
and the development of individuals prone to aggression and violence (e.g., Bandura, 1973; Berkowitz, 1984, 1993; Geen, 2001;
Huesmann, 1988, 1998). The two most detailed current models are
Huesmann’s (1998) script model and Crick and Dodge’s (1994)
social information processing model. Recently, the general aggression model was developed to provide a simplified overview of the
common elements among prior models of the development and
expression of aggressive behavior (Anderson & Bushman, 2002;
Anderson & Carnagey, 2004; Anderson & Huesmann, 2003).
Explicating and comparing these various models lies well outside the scope of this article, but these social– cognitive models
allow several important predictions concerning the likely shortterm and long-term effects of exposure to violent video games. In
general, both short-term and long-term effects of environmental
variables (e.g., insult, physical pain, violent media) on aggressive
behavior operate by affecting cognitive, emotion, and/or arousal
systems.
Aggression Facilitation and Inhibition
Social– cognitive models of aggression distinguish between factors that facilitate the emergence of aggression from those that
inhibit it (e.g., Anderson & Huesmann, 2003; Bandura, Barbaranelli, Caprara, & Pastorelli, 1996; Berkowitz, 1984). Common
facilitating factors in the immediate situational context include
aggression cues (e.g., weapons, violent media) and unpleasant
situational events that put people in a bad mood (e.g., provocation,
frustration, hot temperatures, loud noises, unpleasant odors, pain).
Inhibiting factors include fear of retaliation, negative emotional
reactions to images and thoughts of violence, moral beliefs opposing violence, and pleasant situational events that put people in a
good mood.
Short-Term Versus Long-Term Effects
Short-term effects are those in which a person plays a video
game for a brief time (e.g., 15 min) before relevant measures are
obtained. Usually, short-term effects are assessed in experimental
studies conducted in labs or in schools. Long-term effects are those
that accrue from repeated exposures over a relatively long period
of time, such as months or years. Long-term effects typically are
assessed in cross-sectional and longitudinal studies.
For theoretical reasons, the effects of video game violence might
differ as a function of whether one is discussing short- or longterm effects. This is because the same stimulus can have multiple
effects on several factors that facilitate or inhibit aggression. For
example, playing a video game with sanitized violence versus a
bloody version of the same game may lead to similar levels of
aggressive behavior in the immediate situation, whereas repeated
exposure to one or to the other version may lead the bloody version
to have larger long-term effects. This could happen if both versions
equally prime aggressive scripts while being played but lead to
differential changes in more stable, long-term factors, such as
emotional desensitization to violence, after the game has been
VIOLENT VIDEO GAME EFFECTS
turned off. On the other hand, if the bloody version not only primes
aggressive behavioral scripts but also increases arousal, it might
lead to more aggression in the immediate situation than does the
sanitized version.
Immediate, short-term effects are mainly the result of priming
existing knowledge structures, such as various types of schemata
and scripts (see Bushman & Huesmann, 2006). Priming processes
require only (a) a person who already has at least a few welldeveloped aggression scripts and (b) brief exposure to a video
game that requires violent action. There need be no surface-level
similarity between the violence in the video game and the aggression measure, as long as the person’s aggression scripts have been
activated. That is, the game characters do not need to be similar to
the player or the player’s later real-world target, and the violence
in the game does not need to be similar to the player’s real-world
aggression options. Once aggressive scripts have been activated,
additional exposure to the violent video game is unlikely to have
more than a minimal impact on later aggressive behavior. If
priming of existing knowledge structures is the main process
underlying an observed increase in aggression following video
game play, playing the randomly assigned games for 15 min versus
30 min should make little difference, all else being equal.
Short-term effects might also reflect mimicry or observational
learning of new behaviors and of new beliefs about their likely
success. If the main process underlying an observed short-term
violent video game effect is such mimicry/observational learning,
greater exposure to the violent game (e.g., 30 vs. 15 min) should
lead to better learning of the new aggression script and, in the right
circumstance, to larger increases in aggression. The context most
likely to favor this type of short-term effect is when the participants do not already have well-learned aggression scripts (e.g.,
very young children); when the aggressive behavior being modeled in the game is novel; and when the aggressive behavior test
situation closely resembles the video game in terms of the characters, the provocation, and the possible aggressive action that is
available to the participants. Such conditions are rarely (or never)
encountered in the existing violent video game experiments, which
is why most video game violence researchers believe that the
existing short-term effects are mainly the result of priming effects
(e.g., Anderson et al., 2003, 2007; Bushman & Huesmann, 2006;
Kirsh, 2006; Krahé, 2001).
Long-term effects mainly result from relatively permanent
changes in beliefs, expectancies, scripts, attitudes, and other related person factors that are brought about by repeated exposure to
video game violence. Because these person factors are relatively
stable, repeated exposure to video game violence (or to other
environmental risk factors) is required to create significant change.
Playing a violent video game one time for 20 min will not change
a well-adjusted adolescent into a potential school shooter, with all
of the anger, hostile beliefs, expectations, and personality traits
that go along with such extreme behavior. But repeated exposure
to violent media is expected to lead to measurable changes in the
chronic accessibility of aggression-related knowledge structures
(e.g., aggression scripts, attitudes and beliefs that support aggressive action) and in relatively automatic reactions to scenes or
thoughts of violence (e.g., lack of empathy, physiological desensitization). Another factor important in understanding long-term
effects of exposure to violent media is whether the person’s
environment encourages or discourages aggression. For example,
155
some cultures are relatively supportive of certain types of violence,
whereas other cultures condemn them. Similarly, different families
within a culture may respond differently. This may be why it
appears that having parents who are very involved in one’s media
usage sometimes acts as a protective factor (e.g., Anderson et al.,
2007). Of course, if the highly involved parents actively encourage
violent behavior, they are likely to exacerbate the media violence
effect.
Aggressive Cognition Versus Affect and Arousal
Video games can be exciting, fun, frustrating, exhilarating, and
boring. Being the target of potential harm, even in the virtual world
of video games, is likely to prime aggressive cognitions and
emotions and to increase physiological arousal. The aggressive
cognition aspect is of particular interest for two reasons. First,
many situational factors can increase arousal and anger, even
certain nonviolent video games. For instance, race driving video
games, sports video games, and even perceptual/motor skills
games that require intense concentration and rapid responses (e.g.,
Tetris, Bejeweled) can increase heart rate and blood pressure.
Similarly, video games that are too fast paced or too difficult for
the player are likely to increase frustration and anger, which in turn
might activate aggressive thoughts. But violent video games, by
their nature, require the activation of aggressive thoughts, whereas
nonviolent games do not require it. Second, the repeated activation
of aggressive thoughts, both novel ones (especially in children)
and well-practiced ones, is the most likely route to relatively
permanent changes in the person, because the activation of
aggression-related knowledge structures becomes more automatic
and chronic with repetition and eventually becomes part of the
person’s personality (Strack & Deutsch, 2004; Wegner & Bargh,
1998). The negative affect and physiological arousal instigated by
a video game (violent or nonviolent) likely dissipate fairly quickly
and are less likely to leave long-term traces in the brain than are
the cognitive learning and overlearning of aggression-related perceptual and social schemata (including aggressive behavioral
scripts) that are rehearsed constantly while playing violent games.
Predictions
Before spelling out our specific predictions, we want to raise
two key points. First, predicting the pattern of all the possible
combinations of variables in video game studies requires a thorough knowledge of which processes are engaged by the video
game. Will a third-person shooter have a different impact on
immediate aggression than a first-person shooter? Will gorier
games have a bigger impact than less gory games? Without knowledge of how well each specific game activates aggressive thoughts,
feelings, and physiological arousal, any prediction is risky at best.
Second, the importance of assessing a host of potential short-term
and long-term effects of different types of violent video games
becomes obvious. Nonetheless, a number of broader scale predictions are possible. For example, all else being equal, participants
randomly assigned to play a violent video game should tend to
behave more aggressively for a short period of time afterward than
those randomly assigned to play an equally fun and equally challenging nonviolent video game. We next offer additional predictions, grouped by outcome variable.
156
ANDERSON ET AL.
Aggressive behavior. We expected to find that playing a
violent video game would increase aggressive behavior in a shortterm experimental context, relative to playing a nonviolent video
game that is equally exciting, arousing, and enjoyable. We expected similar effects in long-term contexts, that is, in crosssectional correlation studies and in longitudinal studies. We expected the largest effects in short-term experimental studies and
the smallest effects in longitudinal studies, once sex has been
controlled. This is because experimental studies generally are
better at controlling for effects of extraneous variables that increase the error variance and therefore decrease the effect-size
magnitude.
Aggressive cognition. The predictions for aggressive cognition were the same as for aggressive behavior.
Aggressive affect and physiological arousal. Brief exposure
to violent video games should, on average, increase physiological
arousal and aggressive affect. An important methodological caveat
is warranted, however. Studies based on violent and nonviolent
video games that have been preselected to be equally arousing
obviously are not appropriate tests of the short-term arousal- and
affect-inducing effects of violent video games. Thus, they should
be excluded from the analyses designed to test this specific hypothesis. The same is true when comparison games have been
preselected to create equivalent affective states.
It is less clear what to expect in long-term contexts, but the
temporary nature of moods and of physiological arousal leads us to
expect either very weak long-term effects or none at all. Weak
long-term effects might occur on aggressive affect indirectly
through habitual increases in aggressive thinking or through problems engendered by habitual aggressive behavior. Weak long-term
effects on arousal might occur in young people through changes to
brain regions that control cardiovascular and other arousal-related
functions. Unfortunately, there are no long-term studies of physiological arousal with which to test this hypothesis.
Empathy/desensitization. It is unclear whether playing a violent video game for a brief period of time should have a detectable impact on measures of desensitization to violence or of
empathy for violence victims. Systematic desensitization therapies
suggest that repeated exposures to gory scenes of violence and to
pain and suffering of others will have some impact on a person’s
physiological reactions to new scenes of violence (desensitization)
and on empathetic responses to victims, but such therapies typically take place over a period of days or even weeks. Thus, we
expected brief exposure to a violent video game would have a
relatively small impact on desensitization and empathy. However,
we expected larger effects in long-term studies and in experimental
studies that involve longer desensitization procedures. Unfortunately, there are no long-duration experimental desensitization
studies of violent video game effects.
Prosocial behavior. Social– cognitive models of social behavior suggest that briefly playing a violent video game should
reduce prosocial or helping behavior in the immediate situation.
The temporary increase in aggressive cognition and affect might
be incompatible with, or might interfere with, empathic thoughts
and emotions that frequently underlie helping behavior. Similarly,
short-term desensitization effects could reduce helping victims of
violence in several ways (Bushman & Anderson, 2009; Carnagey,
Anderson, & Bushman, 2007). We therefore expected that video
game violence would produce short-term decreases in some forms
of prosocial behavior.
On the other hand, many violent video games involve the use of
violence to help others, such as saving the princess, one’s teammates, or all of humanity from enemies that need to be killed.
Thus, it is possible that playing certain types of violent video
games might prime a type of “hero” script and thereby lead to an
increased likelihood of certain limited types of helping behavior. No studies covered by our search period tested this hypothesis, though we are aware of several such studies currently in
progress.
We did not expect to find strong long-term decreases in prosocial behavior, because the types of situations that inspire helping
behavior are relatively unlikely to be of the ambiguous kind that
allow spontaneous priming of aggressive thoughts and feelings.
One exception to this latter prediction concerns helping victims of
violence or injury. Because emotional desensitization to injuryrelated cues (expressions of pain, presence of blood) reduces the
perceived need for help by violence victims, repeated exposure to
violent video games should yield long-term declines in this specific type of prosocial behavior (Carnagey et al., 2007). Unfortunately, there have not been enough direct tests of helping victims
of violence, and there are no longitudinal studies testing this
specific hypothesis.
In sum, theory suggests that violent video game effects on
prosocial behavior should be very context specific. However, the
specific contexts used in existing prosocial behavior studies are of
the type that lead to predictions of a significant decrease in
short-term experimental studies and a small effect (or no effect) in
long-term studies.
The Present Meta-Analysis
We undertook the present meta-analysis for four related
reasons. First, the video game violence research literature is
expanding rapidly, with new studies being reported almost
monthly. An updated meta-analysis is badly needed to take into
account the new research. Second, many of the newer studies
are of better methodological quality than some of the earlier
studies (see the meta-analysis in Anderson et al., 2004). With
this larger sample of higher quality studies, one can use stricter
inclusion criteria for the main analyses of potential moderators
and still have a sufficient sample of studies to yield meaningful
results. In essence this larger set of high-quality studies allows
tests of theoretical propositions that could not be tested in prior
years. Third, there is a growing body of research using Japanese
samples, a literature that has gone largely unnoticed by scholars
in the West. This body of research not only adds to the total
body of studies available for an updated meta-analysis but also
allows examination of whether video game violence effects
occur in a low violence society that differs from the West in so
many important ways. Fourth, the larger body of studies in this
domain allows tests of a number of potentially important moderator variables. For example, in experimental studies the video
game violence effect size may differ as function of whether the
violent game is played from a first-person or third-person
perspective.
VIOLENT VIDEO GAME EFFECTS
Method
Literature Search Procedures
Outcome variables. We focused on six outcome variables,
the first five of which have been used in prior meta-analyses. The
outcome variables were physically aggressive behavior, aggressive
cognition, aggressive affect, physiological arousal, prosocial (helping) behavior, and a combined empathy/desensitization variable.
All are described more fully in the Results section.
Western studies literature search. We searched PsycINFO
and MEDLINE for all entries through 2008 using the following
terms: (video* or computer or arcade) and (game*) and (attack* or
fight* or aggress* or violen* or hostil* or ang* or arous* or
prosocial or help* or desens* or empathy). In addition, we
searched the reference sections of prior meta-analytic and narrative
reviews. We included dissertations, book chapters, and unpublished papers.
Japan studies literature search. There is no search engine
comparable to PsycINFO for psychological research in Japan.
Therefore, we searched CiNii (NII Scholarly and Academic Information Navigator) and Magazine Plus (Nichigai Associates, Inc.)
for all entries through 2008 using the following terms: (terebigemu
[TV game] or konpuutaagemu [computer game]) or bideogemu
[video game]).
From these two searches we retrieved over 130 research reports
that contained some potentilly relevant original data, with over 380
effect-size estimates based on over 130,000 participants. As shown in
Table 1, this is a huge increase since the last comprehensive metaanalysis (Anderson et al., 2004) as well as the most recent metaanalyses by Ferguson (2007a, 2007b; Ferguson & Kilburn, 2009).
Outcome Variable Details
Aggressive behavior. High-quality experimental studies typically measure aggressive behavior using noise blasts, electric
shocks, or hot sauce given to an ostensible partner (in the last case,
the partner is known to hate spicy food; for discussions and studies
of validity, see Anderson, Lindsay, & Bushman, 1999; Bushman &
Anderson, 1998; Carlson, Marcus-Newhall, & Miller, 1989;
Giancola & Parrott, 2008). High-quality nonexperimental studies
typically measure aggressive behavior using standardized ques-
Table 1
A Comparison of the Sizes of Recent Meta-Analyses of Violent
Video Game Effects to That of the Current Meta-Analysis
Violent video game studies
Meta-analysis
No. papers
K
N
Anderson et al., 2004
Ferguson, 2007a
Ferguson, 2007b
Ferguson & Kilburn, 2009
Present article
44
24
17
14
136
97
25
21
15
381
16,534
4,205
3,602
Unknown
130,296
Note. It was not possible to derive N for violent video game effects for
Ferguson and Kilburn (2009) because the reported Ns included studies that
had TV and film effects confounded with video game effects. K ⫽ number
of effects sizes; N ⫽ number of participants.
157
tionnaires (e.g., Buss & Perry, 1992), self-reports, peer reports,
teacher reports, or parent reports. Whenever possible, we used
measures of physical aggression, because that is the type of aggression most frequently modeled and rewarded in violent video
games. In many of the nonexperimental studies, the aggression
measure was a composite of physical and verbal aggression.
Aggressive cognition. Aggressive cognition has been assessed in numerous ways. Short-term experimental studies have
used reading reaction time, story completion, word fragment completion, Stroop interference, speed to recognize facial emotions,
and hostile attribution bias measures. Occasionally, more traitlike
measures of aggressive cognition (such as attitudes toward violence) have been used in short-term experimental studies; these are
inappropriate because they measure stable thoughts and beliefs
that develop over a lifetime and should not be influenced by
playing a video game for a few minutes. Nonexperimental studies
have used measures of trait hostility, hostile attribution bias, attitudes toward violence, hypothetical aggression statements, aggression vignettes, implicit association tests, and normative beliefs
about aggression. A few measures, such as variants of the Implicit
Association Test, have been found to be sensitive to short-term
experimental manipulations as well as to reflect longer term attitudes and so have properly been used in both short-term and
long-term studies (Lane, Banaji, Nosek, & Greenwald, 2007).
Aggressive affect. Aggressive affect measures used appropriately in short-term experimental studies include self-report measures of state hostility, state anger, and feelings of revenge. One
experimental study assessed brain function in regions of the brain
known to be affected by anger. Most measures for nonexperimental studies were self-reported trait anger scales.
Physiological arousal. Physiological arousal was assessed
with measures of heart rate, blood pressure, or skin conductance.
Empathy/desensitization. Empathy refers to the degree to
which a person subjectively identifies and commiserates with a
victim and feels emotional distress. Empathy measures are almost
always based on self-report scales in which participants indicate
the extent to which they empathize with, feel sympathy for, or feel
sorry for a particular person or group of people. In high-quality
studies, state measures are used in short-term experimental contexts, whereas trait measures are used in nonexperimental contexts.
The term desensitization has been used to cover a wide range of
measures, including shorter recommended jail terms for persons
convicted of a violent crime to longer latency to intervene in a
violence episode (e.g., Carnagey et al., 2007). Theoretically, however, desensitization refers to a reduction in negative emotional
response to scenes of violence. The best measures of such effects
are negative emotion-related measures, such as heart rate, skin
conductance, or other physiological indicators of emotion-related
reactions to scenes of violence. In the present article, desensitization specifically refers to a reduction in physiological reactivity to
scenes of violence. Most other measures that have been called
desensitization are actually theoretical sequelae of reduced negative emotional reactions.
Empathy and desensitization are similar in that both refer to
automatic emotional reactions to harm befalling someone else.
They differ in directionality and in type of measurement (physiological vs. self-report). We combined these two outcome variables
into one category because of their conceptual similarity and because there were too few studies to warrant separate meta-
158
ANDERSON ET AL.
analyses. We reverse scored the desensitization effects, so that
negative effects indicated that high exposure to video game violence was associated with high desensitization or low empathy. In
other words, theory predicts negative effect sizes.
Prosocial behavior. Experimental studies used donating of
jelly beans or money, helping someone succeed at a task, or
helping a victim of a staged violent episode. Nonexperimental
studies used self- and other reports of helping behavior.
Methodological Criteria Assessment
Many of the effect-size estimates are from high-quality studies
that used well-established and theoretically appropriate measures
or manipulations of exposure to violent video games and wellestablished, theoretically appropriate outcome measures. However,
other studies suffer from one or more serious weaknesses relative
to the specific hypothesis. For example, some experimental studies
used violent and nonviolent video games that were chosen on the
basis of pilot testing because they yielded equal states of arousal;
obviously, such studies do not provide appropriate tests of the
effect of violent content on arousal. Usually, this type of piloting
procedure was done by the original authors to demonstrate that the
selected video games did indeed yield similar arousal states, so that
other hypotheses could be more accurately tested in the main study
(e.g., Anderson et al., 2004; Anderson & Dill, 2000).5
Other studies used weak or inappropriate measures of exposure
to video game violence, such as the amount of time spent playing
any type of video game rather than time spent playing violent
video games. Indeed, many studies report the correlations of both
the time on violent games and the time on all games with physical
aggression in order to test whether the more theoretically appropriate measure yields larger effects (e.g., Anderson & Dill, 2000;
Anderson et al., 2007).
Of course, meta-analysis researchers always face the dilemma of
dealing with studies of widely varying quality and characteristics.
The common solution is to establish a set of methodological
criteria and then exclude studies that fail to meet these criteria. In
some domains this works well, but in more controversial domains
the inclusion/exclusion decisions themselves become the focus of
extended debate, thus decreasing the value of the meta-analysis
itself. We dealt with this issue in multiple ways. First, we divided
studies into two broad categories, those whose methods reflected
the best practices in the manipulation and measurement of theoretically appropriate independent and dependent variables versus
those that did not.6 The main analyses (including the moderator
analyses) were performed on this set of high-quality studies. Second, we contacted the authors of reports that appeared to have
additional unpublished data that could be used to compute effectsize estimates that met the best practice criteria. In this way, we
were able to obtain several best practice effect-size estimates that
were not in the original reports.7 Third, we conducted several types
of sensitivity analyses. As has been done in other recent metaanalyses (e.g., Chida & Hamer, 2008), the average effect was
estimated for each outcome variable on both the full sample and on
the best practice sample. If both types of effect-size estimates
could be computed from the same study, we kept only the one
based on best practices. The full sample analysis reduces the
plausibility of claims of selection bias, because all potentially
relevant studies are included. The full sample may either under-
estimate or overestimate the true effect sizes, because it includes
studies whose methods might artifactually inflate or deflate the
reported effect size. We therefore also report a comparison of best
practice versus other studies. As a final type of sensitivity analysis, we
used the trim and fill procedure to see how much various effect-size
estimates changed as a result of potential selection (publication) bias.
Best practice inclusion criteria. The inclusion criteria for
best practice studies are listed in Table 2, which also gives examples of criteria violations. A more detailed listing of the specific
violations for specific studies is too lengthy for inclusion in this
article but can be downloaded at the following web page: http://
www.psychology.iastate.edu/faculty/caa/abstracts/2010-2014/
NotBestViolations.pdf. Two independent raters examined each
effect and judged whether it met all six criteria. Initial agreement
was over 93%. Discrepancies were examined and discussed with a
third judge until consensus was reached.
Correlated data. For the longitudinal studies we included
both a longitudinal effect size and a cross-sectional effect size. The
latter was the average of the two cross-sectional effects, one
measured at Time 1 and the other at Time 2.
For studies that reported multiple effects on the same conceptual
outcome variable, we took one of two actions. In those cases where
one measure was clearly better than the others, based on theoretical
relevance (e.g., physical aggression is more relevant to violent
video game effects than is verbal aggression), established validity
(e.g., use of a well-validated multiple item measure of trait physical aggression vs. a new single item measure of trait aggression),
or other empirical evidence offered in the study, we used the best
measure. For example, if a study reported two new outcome
measures of aggressive behavior, and only one of them correlated
significantly with a third variable known to be related to physical
aggression (e.g., trait irritability), we used that measure (e.g., Anderson & Dill, 2000, Study 2). In those cases in which there was not a
clear best measure, we used the average effect size (Bartholow,
Bushman, & Sestir, 2006). Note that we also repeated the main
analyses, always using the average effect, and found essentially the
same results. For all analyses, we used fixed effects models, although
random effects models yielded very similar results.
Coding frame: Moderator variables.
All studies. We coded the following information for each effect
size: research design (experimental, cross-sectional, longitudinal); av5
Of course, the quality of that study relative to tests of violent video
game effects on aggressive cognition would be very high.
6
Study quality also varies within each of these two broad categories.
Although one might attempt a more fine-grained, multilevel assessment of
quality, such an attempt would require more studies than presently exist in
this domain.
7
As should be clear, many of the not best practices effects were never
intended by the original authors as tests of the specific hypothesis for
which they earned not best practices status. In many cases, the not best
practices effects were the result of high-quality procedures used to improve
the precision of the main hypothesis test. In other cases, the not best
practices effect was not part of the main study at all but was simply
reported in a correlation matrix that included other variables that were the
main focus of the article. Thus, neither readers nor authors of original
reports should interpret a not best practices listing as a negative judgment
about the author’s methodological skills or the overall value of the study
that included the not best practices effect.
VIOLENT VIDEO GAME EFFECTS
159
Table 2
Inclusion Criteria for Determining Whether a Study Qualifies as a “Best Practices” Study
Criteria
Examples of inclusion criterion violations
1. The compared levels of the independent variable were appropriate
for testing the hypothesis.
2. The independent variable was properly operationalized.
In a short-term experiment, participants in the “nonviolent” condition
played a video game that contained considerable violence.
In a nonexperimental study, total video game play rather than violent
video game play was used as the predictor variable.
Participant retention was substantially lower in one experimental
condition than another, indicating potential self-selection
of participants.
The hypothesis specifies an effect on aggressive behavior, but the
outcome measure assessed behavior directed toward an inanimate
object rather than a person.
A measure of personality trait aggression was used as the measure of
aggressive behavior in a short-term experimental study.
Pre- and postmanipulation scores were averaged but were not
reported separately.
3. The study had sufficient internal validity in all other respects.
4. The outcome measure used was appropriate for testing the
hypothesis.
5. The outcome measure could reasonably be expected to be
influenced by the independent variable if the hypothesis was true.
6. The outcome variable was properly computed.
erage age (when only “college students” was reported, we assigned an
age of 20); culture (East [Japan, China, Singapore] vs. West); and sex
of participants. We also coded a number of other characteristics
specific to a research design and/or an outcome variable.
Experimental studies. For experimental studies we coded the
following features: violent game player’s perspective (first or third
person); violent game player’s role (hero, criminal, neither); violent game targets (human, nonhuman, both); and duration of time
spent playing the assigned video game immediately prior to assessment of the dependent variable. There are no obvious theorybased predictions for these four moderators for most dependent
variables in short-term experimental studies. For example, if the
short-term effect of playing a violent game on aggressive behavior
is a priming phenomenon, playing the game for 30 min is unlikely
to have a greater impact than playing it for 15 min, unless the
content or difficulty changes a lot in the last 15 min.
Many experimental studies of aggressive behavior used some
version of the competitive reaction time (CRT) task. In this task
participants are told that they are competing against another person
on a series of reaction time trials, that the loser of each trial will
receive a punishment immediately after losing the trial, and that
before each trial participants set the punishment level that their
opponents will receive. The punishment settings are used to assess
physical aggression. Original versions used electric shocks (Taylor, 1967), but in more recent years the punishments usually
involve blasts of white noise. The CRT is one of the most widely
used laboratory techniques for measuring physical aggression and
has been shown to have good external validity (Carlson et al.,
1989; Giancola & Chermack, 1998; Giancola & Parrott, 2008).
Thus, for aggressive behavior studies we also coded whether or not
the CRT task was used.
For experimental studies of aggressive cognition, we coded
whether or not the outcome variable was some type of rapid
automatic cognitive response task (e.g., reaction times) or some
type of more thoughtful measure, such as hostile attributional
style.
Nonexperimental studies. Nonexperimental studies included
several different ways of measuring exposure to violent video
games. We created a dichotomous code that distinguished between
studies in which the amount of time spent playing violent games
was specific to each game or game type (and then summed or
averaged across games or game type), and studies that used some
other measures of exposure to violent video games. An example of
the former type of measure is the one reported by Anderson and
Dill (2000). Their video game violence (VGV) exposure measure
has participants list their five most frequently played games. Participants then indicate for each game how frequently they play that
game, how violent the graphics are, and how violent the content is.
The two violence ratings are averaged and then multiplied by the
frequency. This is done for each game listed, and then these five
scores are averaged. We refer to this as the VGV-specific type of
measure. The second type of measure was used in most of the
Japanese studies. Participants rate how frequently they see violent
scenes in the games they play. In some studies, this violent scenes
rating was then multiplied by a measure of how many hours per
week the participant played video games of any type. We refer to
this as the VGV-general measure.
For nonexperimental studies of aggressive behavior we coded
whether the measure was of physical aggression versus a mixture
of physical and some other type (most commonly verbal). We also
coded whether or not the measure was primarily composed of
more extreme physical aggression that is illegal (violence, such as
assault).
For nonexperimental studies of aggressive cognition we coded
whether the measure was of trait hostility versus some other type
of aggressive cognition, such as attitudes toward violence or hostile attributional style. We did this because a large number of
studies used the Hostility subscale of the Aggression Questionnaire (Buss & Perry, 1992) and because that measure is very
similar to trait anger, which is not theoretically expected to yield
strong cross-sectional or longitudinal effects.
For longitudinal studies we coded the length of time between the
initial and the final assessment period. This ranged from 3 to 30
months.
Partial correlations. Normally, partial correlations are not
used in meta-analyses because the statistical theory underlying
meta-analytic procedures assumes that one is working with raw
(zero-order) correlations. This was not an issue for experimental
studies, even when separate effects were not reported by sex,
because random assignment removes any unwanted correlation
between the independent variable and sex.
ANDERSON ET AL.
160
As noted earlier, however, a number of nonexperimental studies
in this domain either ignored sex or reported finding no Sex ⫻
Video Game Violence (VGV) interaction and then reported an
overall VGV effect that combined across sex. Theoretically this
could inflate the VGV effect, because males tend to play more
violent video games and tend to be more physically aggressive
than females. Of course, if violent video games actually do increase physical aggression (and other aggression-related outcome
variables), controlling for sex could lead to artificially low effectsize estimates. Our solution to these issues was to contact researchers with a request for additional data that would allow us either to
get separate estimates for males and females or to statistically
partial out the effect of sex. We then created two overlapping data
sets. One, labeled the “best raw” data set, consisted of all best
practices studies with effects in their rawest form.8 The other,
labeled the “best partials” data set, contained only effects that had
been corrected for sex, either by separate estimates for males and
females or by use of partial correlations.
For longitudinal studies, the best raw data set contained the
correlations between Time 1 VGV exposure and Time 2 outcomes.
For the best partials data set, the effect sizes were partial correlations with both sex and Time 1 outcomes partialed out. The main
analyses and all of the moderator analyses were carried out on the
best partials data set. Thus, results from the best partials data set
are very conservative estimates and may well underestimate the
true video game effects. However, comparing these effects to
the corresponding best raw effects gives another indication of the
strength (or weakness) of the overall effects of violent video
games.
Our final sample consisted of 381 effect-size estimates based on
130,296 participants. Of these, over half (221 effects) met the best
practice inclusion criteria. Table 3 illustrates the breakdown by
outcome variable and, for the best partials data set (the one on
which subsequent analyses focus), by research design and culture.
Appendix A lists the sources that contributed at least one best
practice effect. Appendix B lists additional studies that contributed
effects that did not meet the best practice criteria. (The Appendices
can be found in the online supplemental materials.)
Table 3
Characteristics of the Samples
Full sample
Variable
Category
Aggressive behavior
Aggressive cognition
Aggressive affect
Prosocial behavior
Empathy/desensitization
Physiological arousal
Total
Culture
East
West
Research design
Experimental
Cross-sectional
Longitudinal
Note.
K
N
Best raw
K
N
Best partials
K
N
140 68,313 79 21,681 75 18,751
95 24,534 59 16,271 53 12,598
62 17,370 37 9,191 35 7,543
23
9,645 16 6,906 16 6,905
32
8,528 15 6,580 14 6,268
29
1,906 15
969 15
969
381 130,296 221 61,598 208 53,034
K ⫽ number of effects; N ⫽ total sample size.
64 32,436
144 20,598
92 8,705
82 28,788
34 15,541
About one third of the best practice effects and over half of the
participants were from Eastern cultures, mainly Japan. The majority of the remaining best practice effects were from U.S. samples,
but samples also came from Australia, Germany, Italy, the Netherlands, Portugal, and the United Kingdom.9
Meta-analytic procedures. All effects sizes were converted
to the correlation coefficient, denoted by r. We used the software
program Comprehensive Meta-Analysis and used a fixed effects
model so that we could assess the heterogeneity in various subsets
of studies. For each outcome variable we first computed the overall
average effect size and then computed a moderator analysis based
on type of study. This was done for both the best raw and the best
partials data sets. We then conducted more detailed moderator
analyses on the best partials data.
We also conducted analyses to address the possibility that
results might be affected by selection bias (also called publication
bias) in the sample of studies included in the meta-analyses. The
concern is based on the premise that studies that fail to “work” are
less likely to be published, which might bias the results of a
meta-analysis. To assess the possibility that publication bias affected our results, we used the trim and fill procedure (Duval &
Tweedie, 2000a). Trim and fill is based on the assumption that in
the absence of publication bias, the studies will be distributed
symmetrically about the mean effect size (plotted on the x-axis)
relative to standard error (plotted on the y-axis), because the
sampling error is random. In the presence of publication bias,
studies are expected to be systematically missing in a manner that
can be identified by the trim and fill analysis. In the case of
positive effect data, if low-effect nonsignificant studies are missing, we would expect a gap on the left-bottom quadrant in the plot,
where the nonsignificant studies would have been if we had
located them. If, based on other selection mechanisms, high-effect
studies are selectively missing, the gap would be on the right side
of the mean effect. If asymmetry is detected, trim and fill uses an
iterative procedure to remove the most extreme small studies from
the specified side of the funnel plot, re-computing the effect size at
each iteration until the funnel plot is symmetric about the (new)
effect size. In theory, this will provide an unbiased estimate of the
effect size. Although trimming yields an adjusted effect size, it
also reduces the variance of the effects, resulting in a too-narrow
confidence interval. Therefore, the algorithm then adds the original
studies back into the analysis and imputes a mirror image for each.
This fill has no impact on the point estimate but produces a better
estimate of the variance (Duval and Tweedie, 1998, 2000a,
2000b). The major advantage of this approach is that it addresses
an important question, What is the best estimate of the unbiased
effect size? But, as noted earlier, this estimate should not be
interpreted as the true effect size, because it is based on imputed
data. Furthermore, if there is a true, theoretically meaningful
relationship between effect size and sample size (e.g., researchers
use larger samples when conducting longitudinal studies because
8
For several studies, the only estimates possible were some type of
partial correlation. Rather than discard potentially useful information, we
kept these in both of the best practices analyses.
9
The totals across outcome variables in the table are not all independent
samples, because many studies reported multiple outcome variables for the
same sample.
VIOLENT VIDEO GAME EFFECTS
they know longitudinal effects are smaller than cross-sectional
ones), the trim and fill procedure can erroneously adjust the
average effect sizes. Ideally, the trim and fill procedure is used on
appropriate subsets of studies, and the difference between the
original effect size and the trim and fill adjusted effect size should
be used to get a feel for the possible biasing effect of publication
or selection bias, not as an estimate of the true effect size.
Results
Analysis Plan
Main and moderator analyses were done separately for each
outcome variable. In other words, we conducted six independent
meta-analyses. We first present results of analyses on the best
practice effects, for both the best raw and the best partials data sets.
These are the effects that met the inclusion criteria. We then
present more detailed moderator analyses on the best partials data.
Finally, we present results from several sensitivity analyses for all
outcome variables, including analyses of methodological quality as
a moderator of effect size. Note that in this section, “study” refers
to a sample on which an effect size was computed.
Violent Video Game Effects on Aggressive Behavior
Main analyses. Table 4 presents the results of the main analyses on aggressive behavior for both the best raw and the best
partials data.10 Figure 1 illustrates several main points from these
analyses. First, regardless of research design and regardless of
whether the standard zero-order correlation approach or the much
more conservative partial correlation approach was used, VGV
exposure was significantly related to higher levels of aggressive
behavior. Most notably, in longitudinal studies even when sex and
Time 1 aggressive behavior were controlled, amount of violent
video game play at Time 1 significantly predicted an increase in
aggressive behavior at Time 2.
Second, partialing out sex effects (in cross-sectional and longitudinal studies) and Time 1 aggressive behavior effects (in longitudinal studies) greatly reduced the average effect size of VGV.
Third, in the best partials data, experimental studies yielded the
largest effects whereas longitudinal studies yielded the smallest
effects. Fourth, ignoring research design led to very large heterogeneity effects. Fifth, when research design, sex, and Time 1
effects were controlled, none of the heterogeneity effects were
significant. In sum, this much larger meta-analysis, with over 70
independent effects involving over 18,000 participants from multiple countries, ages, and culture types, yielded strong evidence
that playing violent video games is a significant risk factor for both
short-term and long-term increases in physically aggressive behavior.
Additional moderator tests— best partials data. We conducted additional moderator tests within each research design,
even though the heterogeneity test results yielded little evidence
that the studies within design type came from different populations. There were two reasons for doing these additional tests.
First, there are several specific comparisons that are of special
interest for theoretical (e.g., culture), methodological (e.g., how to
161
measure VGV exposure), or public policy reasons (e.g., player
perspective). Second, the heterogeneity tests are omnibus tests; it
is possible that more focused tests will yield significant differences.
Culture. The effect of culture (Eastern vs. Western) was not
significant in any of the research designs. The average effect in
experimental studies was slightly larger in Eastern than in Western
studies but not significantly so, Q(1) ⫽ 0.28, p ⬎ .50. In crosssectional studies, the VGV effect was slightly larger in Western
than in Eastern studies but not significantly so, Q(1) ⫽ 1.08, p ⬎
.20. In longitudinal studies the VGV effect size was somewhat
larger in Western (r⫹ ⫽ .126, K ⫽ 5, N ⫽ 1,037) than in Eastern
studies (r⫹ ⫽ .059, K ⫽ 7, N ⫽ 3,392). This effect was marginally
significant, Q(1) ⫽ 3.52, p ⬍ .07.
Sex. There was no evidence that the VGV effect on aggressive
behavior differed for males and females ( ps ⬎ .10). The VGV
effect was slightly larger for females in experimental and longitudinal studies and was slightly larger for males in the crosssectional studies, but none of these differences were significant.
Age. Average age of the participants was not significantly
related to the VGV effect sizes in experimental or longitudinal
studies ( ps ⬎ .50). However, it is important to note that there were
no longitudinal studies on participants older than 16. For crosssectional studies there was a marginally significant effect of age
(b ⫽ ⫺.005, Z ⫽ ⫺1.82, p ⬍ .07). Studies with older participants
tended to yield slightly smaller effect sizes than did those with
younger participants.
Moderators specific to experiments. Of the moderators specific to experiments, the only one with at least a marginally
significant effect was the CRT variable, Q(1) ⫽ 2.90, p ⬍ .09.
Experimental studies that used some version of the CRT task
(r⫹ ⫽ .188, K ⫽ 15, N ⫽ 1,724) yielded slightly smaller effects
than did those with some other measure of aggression (r⫹ ⫽ .259,
K ⫽ 12, N ⫽ 789). None of the other moderators associated with
experimental design approached significance (i.e., player perspective, player role, target type, time on game).
Moderators specific to nonexperimental studies. In the crosssectional studies, one of the additional moderators was marginally
significant. Studies with a pure physical aggression measure
(r⫹ ⫽ .184, K ⫽ 28, N ⫽ 7,137) yielded slightly larger VGV
effects than did studies that used some mixed aggression measure
(r⫹ ⫽ .153, K ⫽ 8, N ⫽ 4,672), Q(1) ⫽ 2.82, p ⬍ .10. The method
of measuring VGV exposure did not approach significance ( p ⬎
.20), but the average effect was slightly larger when VGV-specific
measures were used than when VGV-general measures were used.
The violence moderator (violent behavior vs. aggressive but not
violent behavior) also did not approach significance ( p ⬎ .30).
For longitudinal studies, the method of measuring violent video
game exposure was significantly related to the magnitude of the
effect size, Q(1) ⫽ 6.81, p ⬍ .01. Studies that used VGV-specific
type measures (r⫹ ⫽ .152, K ⫽ 4, N ⫽ 902) yielded significantly
larger VGV effects than did those that used VGV-general type
measures (r⫹ ⫽ .055, K ⫽ 8, N ⫽ 3,527). However, this
moderator was somewhat confounded with the East/West mod10
For completeness, we also included results from the full sample, but
these results are not discussed in any detail.
ANDERSON ET AL.
162
Table 4
Aggressive Behavior: Average Effect of Violent Video Game Exposure by Study Design for Best Raw, Best Partials, and Full Sample Data Sets
Test of null
(two-tailed)
Effect size and 95% CI
Design
N
K
Point estimate
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
2,513
4,526
14,642
27
12
40
.210
.203
.262
21,681
79
2,513
4,429
11,809
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
Note.
LL
UL
Heterogeneity
z
p
Q
df (Q)
p
Best raw
0.172
0.248
0.175
0.231
0.247
0.277
10.512
13.787
32.291
.000
.000
.000
.244
0.231
0.256
36.422
.000
19.41
40.72
207.99
268.12
16.74
284.86
26
11
39
76
2
78
.819
.000
.000
.000
.000
.000
27
12
36
.210
.075
.171
Best partials
0.172
0.248
0.045
0.104
0.154
0.189
10.512
4.974
18.732
.000
.000
.000
18,751
75
.154
0.140
0.168
21.118
.000
19.41
9.22
49.55
78.19
40.17
118.36
26
11
35
72
2
74
.819
.601
.052
.289
.000
.001
3,464
5,513
59,336
45
14
81
.181
.198
.189
Full sample
0.148
0.213
0.172
0.223
0.181
0.196
10.538
14.812
46.412
.000
.000
.000
68,313
140
.189
0.182
49.838
.000
79.08
41.53
771.85
892.46
0.70
893.16
44
13
80
137
2
139
.001
.000
.000
.000
.706
.000
0.196
Effect sizes measured as r. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
erator, and this clouded interpretation. Three of the four VGVspecific studies were from the West, but only two of the eight
VGV-general studies were from the West. Both types of measures, however, yielded significant longitudinal effects ( ps ⬍
Figure 1. Effects of playing violent video games on aggressive behavior:
Averages and 95% confidence intervals by research design. Exp ⫽ experimental studies (same in best raw and best partials data); CrSec ⫽ crosssectional studies; Raw ⫽ data from best raw samples; SA ⫽ sex adjusted
(data from best partials samples); Long ⫽ longitudinal studies; VGV
Specific ⫽ studies that used the more specific type of video game violence
exposure measure; T1 & SA ⫽ Time 1 and sex adjusted.
.001 and .005, respectively). Neither the time between measurements nor the physical versus mixed aggression moderators
approached significance.
Summary of main findings. Regardless of research design or
conservativeness of analysis, exposure to violent video games was
significantly related to higher levels of aggressive behavior. For
experimental studies, r⫹ ⫽ .210. For cross-sectional studies the
best raw and best partials analyses yielded average effect sizes of
r⫹ ⫽ .262 and .171, respectively. For longitudinal studies the best
raw and best partials analyses yielded average effect sizes of r⫹ ⫽
.203 and .075, respectively. Neither culture nor sex yielded any
significant moderator effects.
The fact that significant positive effect sizes were obtained in
short- and long-term contexts confirms our main theoretical hypotheses. Of particular importance is the finding of a significant
longitudinal effect. This shows that playing violent video games
can increase aggression over time. Thus, the present findings,
especially the longitudinal ones, fill the main gap in the empirical
literature on violent video game effects. Furthermore, these effects
appear to generalize across culture.
The marginally significant age effect suggests that children
might be more susceptible than young adults to violent video game
effects, but more research specifically targeted to this question is
needed. The lack of a time on game effect in experimental studies
suggests that these effects are largely based on priming processes
that are triggered in the first few minutes of game play. Experimental studies that vary time on game within the same study are
needed to provide a more precise look at this question.
VIOLENT VIDEO GAME EFFECTS
Violent Video Game Effects on Aggressive Cognition
Main analyses. Table 5 presents the results of the main
analyses on aggressive cognition. As with aggressive behavior,
VGV exposure was significantly related to increased levels of
aggressive cognition, regardless of research design and regardless of whether zero-order correlations or the more conservative
partial correlation approach was used. Furthermore, even when
sex and Time 1 aggressive cognition were controlled, amount of
violent video game play at Time 1 predicted a significant
increase in aggressive cognition at Time 2.
Also as expected, partialing out sex and Time 1 aggressive
cognition effects reduced the average VGV effect size in nonexperimental studies. In the best partials data, experimental studies
yielded the largest effects, whereas longitudinal studies yielded the
smallest effects. Once again, ignoring research design led to very
large heterogeneity effects. Finally, when research design, sex, and
Time 1 effects were controlled, none of the heterogeneity effects
were significant. In sum, this much larger meta-analysis, with over
50 independent effects involving over 12,000 participants from
multiple countries, ages, and cultures, yielded strong evidence that
playing violent video games increases aggressive cognition in both
short- and long-term contexts.
Additional moderator tests— best partials data.
Culture. Culture (Eastern vs. Western) was not a significant
moderator in either the experimental or the cross-sectional studies.
However, in longitudinal studies the VGV effect was significantly
163
larger in Western (r⫹ ⫽ .137, K ⫽ 3, N ⫽ 710) than in Eastern
studies (r⫹ ⫽ .038, K ⫽ 5, N ⫽ 2,602), Q(1) ⫽ 5.50, p ⬍ .02.
Sex. There were insufficient data to test the sex moderator
effect in experimental and longitudinal studies. In cross-sectional
studies, the VGV effect was slightly larger for females but not
significantly so.
Age. There were too few longitudinal studies to test the moderating effect of age. Age was not a significant moderator in either
experimental or cross-sectional studies.
Moderators specific to experiments. None of the moderators
specific to experiments (i.e., player perspective, player role, target
type, time on game, type of aggressive cognition measure) approached significance.
Moderators specific to nonexperimental studies. For crosssectional studies neither of the additional moderators approached
significance (VGV exposure measure, type of aggressive cognition
measure). However, for longitudinal studies, the VGV measure of
video game exposure moderator was marginally significant,
Q(1) ⫽ 3.52, p ⬍ .07. Studies that used VGV-specific measures
(r⫹ ⫽ .113, K ⫽ 4, N ⫽ 891) yielded larger VGV effects than did
those that used VGV-general type measures (r⫹ ⫽ .040, K ⫽ 4,
N ⫽ 2,421). Also of importance was the finding that type of
aggressive cognition measure (trait hostility vs. other) was completely confounded with culture and therefore yielded exactly the
same moderation effect as reported earlier for culture. The five
longitudinal studies from Japan all used a trait hostility measure,
whereas the three studies from Western cultures used hostile
Table 5
Aggressive Cognition: Average Effect of Violent Video Game Exposure by Study Design for Best Raw, Best Partials, and Full Sample
Data Sets
Test of null
(two-tailed)
Effect size and 95% CI
Design
N
K
Point estimate
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
2,887
3,408
9,976
24
8
27
.217
.115
.183
16,271
59
2,887
3,312
6,399
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
LL
UL
Heterogeneity
z
p
Q
df (Q)
p
Best raw
0.181
0.252
0.082
0.148
0.164
0.202
11.695
6.728
18.445
.000
.000
.000
.175
0.160
0.190
22.440
.000
35.11
13.08
185.56
233.75
18.74
252.49
23
7
26
56
2
58
.051
.070
.000
.000
.000
.000
24
8
21
.217
.059
.114
Best partials
0.181
0.252
0.025
0.093
0.090
0.138
11.695
3.400
9.128
.000
.001
.000
12,598
53
.123
0.106
0.141
13.826
.000
35.11
7.81
19.84
62.76
40.49
103.25
23
7
20
50
2
52
.051
.349
.468
.106
.000
.000
4,289.5
4,178.5
16,066
48
9
38
.207
.110
.164
Full sample
0.177
0.236
0.080
0.140
0.149
0.179
13.496
7.142
20.951
.000
.000
.000
24,534
95
.162
0.150
25.528
.000
90.00
13.50
269.06
372.56
20.41
392.98
47
8
37
92
2
94
.000
.096
.000
.000
.000
.000
0.175
Note. Effect sizes measured as r. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
ANDERSON ET AL.
164
attribution bias or attitude/belief measures. Thus, it is impossible
to know whether this moderation effect results from culture differences in the VGV longitudinal effect or from measurement
instrument differences.
Summary. Exposure to violent video games was significantly
related to higher levels of aggressive cognition, regardless of
research design or conservativeness of analysis. For experimental
studies, r⫹ ⫽ .217. For cross-sectional studies the best raw and
best partials analyses yielded average effect sizes of r⫹ ⫽ .183
and .114, respectively. For longitudinal studies, the best raw and
best partials analyses yielded average effect sizes of r⫹ ⫽ .115
and .059, respectively. Culture significantly moderated the longitudinal VGV effect, but the VGV effect was significantly greater
than zero in both cases. Furthermore, in this small set of longitudinal studies culture was perfectly confounded with type of cognition measure. In addition, studies that used a VGV-specific
measure yielded larger effects than those that used a VGV-general
measure. Three of the four VGV-specific studies and none of the
VGV-general studies were from the West. Therefore, it is unclear
whether the smaller longitudinal effect in studies from Japan was
the result of culture or of either or both of two measurement
instrument differences. Additional research could easily resolve
this.
As with the aggressive behavior results, perhaps the most important finding relative to prior meta-analyses is the significant
longitudinal effect of VGV on aggressive cognition. In combination with the experimental and the cross-sectional findings, the
data provide strong evidence that playing violent video games is a
significant causal risk factor for both short- and long-term increases in aggressive thinking.
Violent Video Game Effects on Aggressive Affect
Main analyses. Table 6 presents the results of the main analyses on aggressive affect. As with aggressive behavior and aggressive cognition, VGV exposure was significantly related to higher
levels of aggressive affect regardless of research design and regardless of whether zero-order correlations or the more conservative partial correlations were used. Furthermore, even when sex
and Time 1 aggressive affect were controlled, amount of violent
video game play at Time 1 predicted a significant increase in
aggressive affect at Time 2.
As with aggressive behavior and aggressive cognition, research
design was a significant moderator of the VGV effect on aggressive affect. Experimental studies yielded the largest effects and
longitudinal studies the smallest.
Additional moderator tests— best partials data. There were
too few longitudinal studies for us to do any additional moderator
analyses. Furthermore, none of the moderator variables yielded a
significant effect in experimental or cross-sectional studies, even
though there was evidence of significant heterogeneity within
experimental studies.
Violent Video Game Effects on Prosocial Behavior
Main analyses. Table 7 presents the main results on prosocial
behavior. VGV exposure was significantly related to lower levels
Table 6
Aggressive Affect: Average Effect of Violent Video Game Exposure by Study Design for Best Raw, Best Partials, and Full Sample
Data Sets
Test of null
(two-tailed)
Effect size and 95% CI
Design
N
K
Point estimate
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
1,454
2,602
5,135
21
5
11
.294
.075
.101
9,191
37
1,454
2,602
3,487
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
Note.
LL
UL
Heterogeneity
z
p
Q
df (Q)
p
Best raw
0.245
0.341
0.037
0.113
0.074
0.128
11.289
3.836
7.227
.000
.000
.000
.124
0.104
0.144
11.883
.000
49.15
13.19
16.56
78.91
53.18
132.08
20
4
10
34
2
36
.000
.010
.085
.000
.000
.000
21
5
9
.294
.039
.110
Best partials
0.245
0.341
0.0001
0.077
0.077
0.143
11.289
1.967
6.509
.000
.049
.000
7,543
35
.121
0.098
0.143
10.481
.000
49.15
9.76
9.78
68.70
63.83
132.53
20
4
8
32
2
34
.000
.045
.281
.000
.000
.000
3,015
3,373
10,982
37
6
19
.181
.082
.145
Full sample
0.146
0.216
0.048
0.116
0.126
0.163
9.863
4.768
15.215
.000
.000
.000
17,370
62
.139
0.124
18.293
.000
111.22
13.73
153.17
278.12
16.87
294.99
36
5
18
59
2
61
.000
.017
.000
.000
.000
.000
0.153
Effect sizes measured as r. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
VIOLENT VIDEO GAME EFFECTS
165
Table 7
Prosocial Behavior: Average Effect of Violent Video Game Exposure by Study Design for Best Raw, Best Partials, and Full Sample
Data Sets
Test of null
(two-tailed)
Effect size and 95% CI
Design
N
K
Point estimate
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
633
2,778
3,495
4
5
7
⫺.182
⫺.114
⫺.093
6,906
16
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
633
2,777
3,495
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
LL
UL
Heterogeneity
z
p
Q
df (Q)
p
Best raw
⫺0.257
⫺0.106
⫺0.151
⫺0.077
⫺0.126
⫺0.060
⫺4.599
⫺6.022
⫺5.506
.000
.000
.000
⫺.110
⫺0.133
⫺0.086
⫺9.125
.000
3.79
15.91
19.07
38.77
4.46
43.23
3
4
6
13
2
15
.285
.003
.004
.000
.107
.000
4
5
7
⫺.182
⫺.062
⫺.094
Best partials
⫺0.257
⫺0.106
⫺0.099
⫺0.025
⫺0.127
⫺0.061
⫺4.599
⫺3.268
⫺5.544
.000
.001
.000
6,905
16
⫺.089
⫺0.113
⫺0.066
⫺7.404
.000
3.79
2.32
19.25
25.36
7.74
33.10
3
4
6
13
2
15
.285
.677
.004
.021
.021
.004
875
2,778
5,992
8
5
10
⫺.161
⫺.114
⫺.086
Full sample
⫺0.226
⫺0.095
⫺0.151
⫺0.077
⫺0.111
⫺0.061
⫺4.748
⫺6.022
⫺6.659
.000
.000
.000
9,645
23
⫺.101
⫺0.121
⫺9.904
.000
5.26
15.91
29.29
50.46
5.05
55.51
7
4
9
20
2
22
.629
.003
.001
.000
.080
.000
⫺0.081
Note. Effect sizes measured as r. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
of prosocial behavior regardless of research design and regardless
of whether zero-order or the partial correlations were used. Even
when sex and Time 1 prosocial behavior were controlled, amount
of violent video game play at Time 1 predicted a significant
decrease in prosocial behavior at Time 2 in longitudinal studies.
Yet again, research design was a significant moderator of the
VGV effect on prosocial behavior, with experimental studies
yielding the largest (negative) effects and longitudinal studies the
smallest. Once sex and Time 1 effects were partialed out, there was
no evidence of heterogeneity in the experimental or longitudinal
effects, but there was in the cross-sectional studies.
Additional moderator tests— best partials data. There were
too few experimental and longitudinal studies to do any additional
moderator analyses on them. For cross-sectional studies we were able
to test for culture, sex, and VGV type of measure effects. On average,
the VGV effect on prosocial behavior was larger in the Western
studies (r⫹ ⫽ ⫺.225, K ⫽ 2, N ⫽ 347) than in Eastern studies (r⫹
⫽ ⫺.079, K ⫽ 5, N ⫽ 3,148), Q(1) ⫽ 6.83, p ⬍ .01. There also was
a significant effect of violent video game exposure measure, Q(1) ⫽
13.69, p ⬍ .001. Studies that used the VGV-specific type of measure
yielded larger (negative) effects (r⫹ ⫽ ⫺.186, K ⫽ 3, N ⫽ 1,074)
than those that used general measures (r⫹ ⫽ ⫺.052, K ⫽ 4, N ⫽
2421). Furthermore, culture and VGV measure were confounded; two
of the three VGV-specific studies came from a Western culture,
whereas all of the VGV-general studies came from Eastern cultures.
Finally, there was no evidence of sex differences in the effect of
violent games on prosocial behavior.
Violent Video Game Effects on
Empathy/Desensitization
Main analyses. Table 8 presents the main results on empathy/
desensitization. VGV exposure was significantly related to less empathy (and more desensitization) regardless of research design and
regardless of whether the zero-order or partial correlations were used.
When sex and Time 1 effects were controlled, research design
was a significant moderator variable. Of course, because there was
only one experimental study, comparisons across designs should
be made with caution.
Additional moderator tests— best partials data. There were
too few experimental and longitudinal studies to do any additional
moderator analyses. For cross-sectional studies, we were able to
test the moderating effects of culture and video game exposure
measure. On average, effect sizes were larger in Western studies
(r⫹ ⫽ ⫺.294, K ⫽ 4, N ⫽ 450) than in Eastern studies (r⫹ ⫽
⫺.144, K ⫽ 5, N ⫽ 3,148), Q(1) ⫽ 9.53, p ⬍ .01. There also was
a significant effect for video game exposure measure, Q(1) ⫽ 4.36,
p ⬍ .05. Studies that used the VGV-specific type of measure
yielded larger (negative) effects (r⫹ ⫽ ⫺.211, K ⫽ 5, N ⫽ 1,177)
than did those using the VGV-general measure (r⫹ ⫽ ⫺.139, K ⫽
4, N ⫽ 2,421). Furthermore, culture and video game exposure
measure were confounded; four of the five VGV-specific studies
came from a Western culture, whereas all of the VGV-general
studies came from Eastern cultures.
ANDERSON ET AL.
166
Table 8
Empathy/Desensitization: Average Effect of Violent Video Game Exposure by Study Design for Best Raw, Best Partials, and Full
Sample Data Sets
Test of null
(two-tailed)
Effect size and 95% CI
Design
N
K
Point estimate
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
249
2,421
3,910
1
4
10
⫺0.138
⫺0.184
⫺0.203
6,580
15
249
2,421
3,598
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
Experimental
Longitudinal
Cross-sectional
Total within
Total between
Overall
LL
UL
Heterogeneity
z
p
Q
df (Q)
p
Best raw
⫺0.258
⫺0.014
⫺0.223
⫺0.145
⫺0.233
⫺0.173
⫺2.175
⫺9.147
⫺12.845
.030
.000
.000
⫺0.194
⫺0.217
⫺0.170
⫺15.873
.000
0.00
24.88
24.46
49.34
1.44
50.78
0
3
9
12
2
14
1.000
0.000
0.004
0.000
0.488
0.000
1
4
9
⫺0.138
⫺0.070
⫺0.163
Best partials
⫺0.258
⫺0.014
⫺0.109
⫺0.030
⫺0.195
⫺0.131
⫺2.175
⫺3.427
⫺9.817
.030
.001
.000
6,268
14
⫺0.126
⫺0.150
⫺0.102
⫺9.999
.000
0.00
12.82
23.10
35.92
12.87
48.79
0
3
8
11
2
13
1.000
0.005
0.003
0.000
0.002
0.000
537
2,796
5,195
11
6
15
⫺0.148
⫺0.160
⫺0.188
Full sample
⫺0.232
⫺0.062
⫺0.196
⫺0.123
⫺0.214
⫺0.162
⫺3.351
⫺8.501
⫺13.671
.001
.000
.000
8,528
32
⫺0.177
⫺0.197
⫺16.383
.000
24.20
37.23
55.96
117.39
2.00
119.38
10
5
14
29
2
31
0.007
0.000
0.000
0.000
0.369
0.000
⫺0.156
Note. Effect sizes measured as r. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
Violent Video Game Effects on Physiological Arousal
Main analyses. All of the physiological arousal studies were
experiments. Overall, playing a violent video game increased
physiological arousal (r⫹ ⫽ .184, p ⬍ .001, K ⫽ 15, N ⫽ 969).
The heterogeneity test was significant, Q(14) ⫽ 30.43, p ⬍ .01.
We were able to conduct moderator analyses for player perspective, player role, game violence target, average age, and gameplaying time. However, none of the moderator tests approached
statistical significance.
Sensitivity Analyses
Full sample. Recall that numerous studies did not meet the
inclusion criteria listed in Table 2. What happens if these methodologically weak studies are included in the main analyses? Are there
systematic differences in average effect size? Table 9 displays the
results. Three main points emerge. First, for each of the six outcome
variables, the violent video game effect size was still significant even
when methodologically weaker studies were included ( ps ⬍ .001).
Second, for five of the outcome variables the methodologically weak
studies yielded smaller effect sizes than did the methodologically
strong studies; in four cases the difference was statistically significant.
The exception was aggressive affect, for which the methodologically
weak studies yielded a significantly larger average effect size. Third,
for each outcome variable even the methodologically weak studies
yielded a significant overall effect.
Trim and fill analyses. Table 10 presents the results of the
trim and fill analyses, as a further check on possible selection/
publication bias. For each outcome variable, we first applied the
trim and fill procedure to the full sample (which included both
methodologically weak and strong studies) and the best raw sample, ignoring research design.
Next, because research design was such a strong and consistent
moderator variable for the best partials samples and because these
samples are the main focus of this article, we also applied the trim
and fill procedure to the best partials data by research design.
Recall that sample size, research design, culture, and specific
research instruments are somewhat confounded in these studies.
Therefore, within each outcome and research design we broke
down the studies into smaller subgroups using the following decision rule: If there was a significant moderator effect, separate
trim and fill analyses were done on the different levels of that
moderator; otherwise, separate analyses were done by culture. Of
course, in several cases there were too few studies for these
breakdowns, in which case the trim and fill procedure was applied
at a higher level of studies. For example, all of the best practices
physiological studies used an experimental design, and only one of
these studies was from an Eastern culture.
The summarized results at the bottom of Table 10 suggest that
if there has been selection or publication bias in favor of theoretical hypotheses in these samples, the bias has been weak and has
had relatively little impact on average effect-size estimates. For
example, experimental studies were weakened by .017 by the trim
and fill imputation procedure. The cross-sectional studies were
weakened by an even smaller amount, .005. Conversely, the longitudinal studies were strengthened by an average of .008. In sum,
VIOLENT VIDEO GAME EFFECTS
167
Table 9
Effects of Methodologically Weak Versus Strong Studies
Test of null
(two-tailed)
Effect size and 95% CI
Methodological
quality
N
K
Point estimate
Weak
Strong
Total within
Total between
Overall
46,632
21,681
61
79
0.163
0.244
68,313
140
0.189
Weak
Strong
Total within
Total between
Overall
8,263
16,271
36
59
0.138
0.175
24,534
95
0.162
Weak
Strong
Total within
Total between
Overall
8,179
9,191
25
37
0.155
0.124
17,370
62
0.139
Weak
Strong
Total within
Total between
Overall
2,739
6,906
7
16
9,645
Weak
Strong
Total within
Total between
Overall
Weak
Strong
Total within
Total between
Overall
LL
UL
p
Q
df (Q)
p
35.506
36.422
.000
.000
0.196
49.838
.000
504.84
284.86
789.70
103.46
893.16
60
78
138
1
139
.000
.000
.000
.000
.000
Aggressive cognition
0.116
0.159
0.160
0.190
12.497
22.440
.000
.000
0.175
25.528
.000
132.46
252.49
384.95
8.02
392.98
35
58
93
1
94
.000
.000
.000
.005
.000
Aggressive affect
0.134
0.176
0.104
0.144
14.060
11.883
.000
.000
0.153
18.293
.000
158.67
132.08
290.75
4.24
294.99
24
36
60
1
61
.000
.000
.000
.039
.000
⫺0.078
⫺0.110
Prosocial behavior
⫺0.116
⫺0.041
⫺0.133
⫺0.086
⫺4.095
⫺9.125
.000
.000
23
⫺0.101
⫺0.121
⫺0.081
⫺9.904
.000
10.33
43.23
53.56
1.95
55.51
6
15
21
1
22
.111
.000
.000
.163
.000
1,948
6,580
17
15
⫺0.116
⫺0.194
Empathy/desensitization
⫺0.160
⫺0.071
⫺0.217
⫺0.170
⫺5.078
⫺15.873
.000
.000
8,528
32
⫺0.177
⫺0.156
⫺16.383
.000
59.26
50.78
110.03
9.35
119.38
16
14
30
1
31
.000
.000
.000
.002
.000
937
969
14
15
0.085
0.184
Physiological arousal
0.020
0.150
0.121
0.245
2.552
5.647
.011
.000
1,906
29
0.135
5.814
.000
10.47
30.43
40.89
4.59
45.48
13
14
27
1
28
.655
.007
.042
.032
.020
Aggressive behavior
0.154
0.172
0.231
0.256
0.182
0.150
0.124
⫺0.197
0.090
0.180
z
Heterogeneity
Note. Effect sizes measured as r. CI ⫽ confidence interval; LL ⫽ lower limit; UL ⫽ upper limit.
there is no evidence that publication or selection bias had an
important influence on the results.
Discussion
Main Findings
Although the meta-analyses in this article revealed numerous
findings about the short- and long-term effects of playing violent
video games, six findings are particularly important. First, social–
cognitive models and other theoretical considerations predicted the
broad pattern of results quite well. As expected, VGV exposure
was positively associated with aggressive behavior, aggressive
cognition, and aggressive affect. These effects were statistically
reliable in experimental, cross-sectional, and longitudinal studies,
even when unusually conservative statistical procedures were
used. Also as expected, VGV exposure was related to desensitization and lack of empathy and to lack of prosocial behavior.
Furthermore, the relative magnitudes of effects for different outcome variables and moderator variables were mainly consistent
with theory. For example, the longitudinal effect of VGV was
somewhat smaller on aggressive affect than on aggressive cognition or behavior. This suggests that the processes underlying media
violence effects are well understood (cf. Anderson et al., 2003).
Second, the VGV effects are significant in Eastern as well as
Western cultures. There are hints that VGV effects may be larger
in Western than Eastern cultures, but these occur only in nonex-
ANDERSON ET AL.
168
Table 10
Results of the Trim and Fill Selection/Publication Bias Analysis
Imputed studies
Sample
Obs. r⫹
Adj. r⫹
Change
Strength
change
Aggressive behavior
Right
Right
0.189
0.244
0.192
0.244
0.003
0.000
0.003
0.000
0
8
Left
0.245
0.207
0.245
0.178
0.000
⫺0.029
0.000
⫺0.029
1
3
Left
Left
0.163
0.182
0.160
0.176
⫺0.003
⫺0.006
⫺0.003
⫺0.006
1
1
Left
Left
0.152
0.055
0.143
0.054
⫺0.009
⫺0.001
⫺0.009
⫺0.001
Left
0.163
0.175
0.217
0.170
0.175
0.199
0.008
0.000
⫺0.018
0.008
0.000
⫺0.018
1
4
Left
Left
0.102
0.127
0.100
0.106
⫺0.002
⫺0.021
⫺0.002
⫺0.021
0
2
Right
0.038
0.137
0.038
0.182
0.000
0.045
0.000
0.045
0.139
0.124
0.294
0.100
0.102
0.294
⫺0.039
⫺0.022
0.000
⫺0.039
⫺0.022
0.000
0.097
0.150
0.039
0.062
0.137
0.049
⫺0.035
⫺0.013
0.011
⫺0.035
⫺0.013
0.011
Prosocial behavior
Right
Right
Right
⫺0.101
⫺0.110
⫺0.182
⫺0.064
⫺0.089
⫺0.125
0.037
0.021
0.057
⫺0.037
⫺0.021
⫺0.057
Right
Right
Left
⫺0.186
⫺0.052
⫺0.062
⫺0.168
⫺0.033
⫺0.070
0.018
0.019
⫺0.008
⫺0.018
⫺0.019
0.008
⫺0.177
⫺0.194
⫺0.179
⫺0.194
⫺0.002
0.000
0.002
0.000
⫺0.144
⫺0.294
⫺0.211
⫺0.139
⫺0.070
⫺0.167
⫺0.294
⫺0.211
⫺0.171
⫺0.070
⫺0.023
0.000
0.000
⫺0.032
0.000
0.023
0.000
0.000
0.032
0.000
0.135
0.184
0.135
0.184
0.000
0.000
0.000
0.000
K
N
Direction
Full
Best raw
Best partials: Experimental
East
West
Best partials: Cross-sectional
East
West
Best partials: Longitudinal
VGV-specific
VGV-general
140
79
27
5
22
36
8
28
12
4
8
10
1
Full
Best raw
Best partials: Experimental
Best partials: Cross-sectional
East
West
Best partials: Longitudinal
East
West
95
59
24
21
6
15
8
5
3
12
0
5
Full
Best raw
Best partials: Experimental
Best partials: Cross-sectional
East
West
Best partials: Longitudinal
62
37
21
9
5
4
5
17
15
0
2
2
1
Left
Left
Right
Full
Best raw
Best partials: Experimental
Best partials: Cross-sectional
VGV-specific
VGV-general
Best partials: Longitudinal
23
16
4
7
3
4
5
10
4
2
2
2
2
Full
Best raw
Best partials: Cross-sectional
East
West
VGV-specific
VGV-general
Best partials: Longitudinal
32
15
9
5
4
5
4
4
2
0
2
0
0
2
0
Full
Best raw and partials
29
15
0
0
Aggressive cognition
Right
Aggressive affect
Left
Left
Empathy/desensitization
Left
Left
Left
Physiological arousal
Overall average
Full average
Best raw average
⫺0.006
⫺0.011
⫺0.007
(table continues)
VIOLENT VIDEO GAME EFFECTS
169
Table 10 (continued)
Imputed studies
Sample
K
N
Direction
Obs. r⫹
Adj. r⫹
Change
Strength
change
⫺0.017
⫺0.017
0.008
Best partials: Experimental average
Best partials: Cross-sectional average
Best partials: Longitudinal average
Note. Strength change is the difference between the observed and adjusted, taking into account the hypothesized effect direction. VGV ⫽ video game
violence; Obs. r⫹ ⫽ observed average effect size; Adj. r⫹ ⫽ Adjusted average effect size.
perimental studies. Indeed, in experimental studies all three outcome variables for which there were sufficient studies (aggressive
behavior, aggressive cognition, and aggressive affect) yielded
slightly larger effects in Eastern studies, though none approached
statistical significance. In the few nonexperimental cases in which
culture yielded a significant moderator effect, it was unclear
whether the difference should be attributed to cultural differences
in vulnerability or to the use of different measures.
Third, there is evidence that the magnitude of the effect size
obtained in a study is influenced by the types of measures one uses.
How one measures exposure to violent video games in nonexperimental contexts influences the magnitude of effects. Measurement
procedures, similar to those used by Anderson and Dill (2000), that
elicit violence and time ratings for each specific game played by
the participant appear to yield larger effect sizes than do other
methods. However, to some extent this finding is confounded with
culture. What is needed to clarify this issue are studies in which
multiple methods of assessing violent video game exposure are
used in the same sample of participants, so that they can be directly
compared. Similarly, it appears that trait hostility measures may
not be the best way to assess aggressive cognition.
Fourth, and perhaps most important, the newly available longitudinal studies provide further confirmation that playing violent
video games is a causal risk factor for long-term harmful outcomes. This is especially clear for aggressive behavior, aggressive
cognition, and empathy/desensitization. But significant longitudinal VGV effects also were found for aggressive affect and for
prosocial behavior.
Fifth, the failures of several specific moderators to yield significant effects are also important for theoretical and practical reasons. In experimental studies, the lack of player perspective (first
or third person), player role (hero, criminal), time on game, and
target (human, nonhuman) all suggest that the short-term effects of
violent video games on aggression are largely the result of priming
processes (see Bushman & Huesmann, 2006). The fact that the
CRT task yielded effect sizes similar to those from other measures
of aggressive behavior in experimental studies strongly contradicts
recent claims that it is only CRT studies that find such effects;
indeed, the average CRT effect was slightly (nonsignificantly)
smaller than for the other measures. The fact that studies using
violent behavior measures did not yield significantly smaller effects than comparable studies using less extreme aggression measures, and that the violent behavior studies by themselves yielded
a significant VGV effect, is important. The facts that sex did not
significantly moderate the findings and that age yielded only one
marginally significant moderation effect suggest that large portions of the population (at least through college age) are susceptible to the harmful effects of violent video game play.
Sixth, our results confirm that partialing out sex effects and (in
longitudinal studies) Time 1 outcome variable effects reduces
average effect size. This is not surprising, of course, but it does
lead to interesting questions about what is the “best” estimate of
the true effect sizes. Partialing out sex effects certainly is a safe,
conservative procedure in that it guarantees that the estimated
average effect size will not be upwardly biased. However, if
violent video game exposure truly plays a causal role in aggression
and in the other outcomes assessed in these meta-analyses (and our
conservative procedures do demonstrate this), partialing out sex
effects yields underestimates of the true effect size. Perhaps the
best solution is to view sex-controlled estimates as the lower
boundary of the true effects and the sex-ignored effects as the
upper boundary.11
Two additional findings warrant highlighting, especially in view
of all the attention that has paid recently to claims of publication
bias and other weaknesses in the violent video game literature.
First, there was no evidence that publication (or study selection
bias) is responsible for the observed relations between exposure to
violent video games and aggressive behavior or the other five
outcome variables. Differences between the observed average effect sizes and the corresponding trim and fill adjusted averages
were small in magnitude and did not substantially change any
overall effects or conclusions. Second, our other sensitivity analyses (see Table 9) showed that even including methodologically
weak studies (studies that would not be analyzed in noncontroversial fields of study) yields relatively small changes in average
effect sizes. Even the weak studies yielded significant overall
effects on all six outcome variables. In sum, the only way one can
“demonstrate” that the existing literature on violent video game
effects does not show multiple causal harmful effects is to use an
incredibly small subset of the existing literature, include some of
the methodologically poorest studies, exclude many of the methodologically strongest studies, and misuse standard meta-analytic
techniques.
Finally, it is important to note that our main analyses were
conducted on the most conservative sample of studies, in which all
explained variance that is confounded with sex or Time 1 outcome
measures is statistically removed from the estimates of the VGV
effects. Thus, the finding of significant VGV effects on all outcome measures and in all research designs speaks to the power of
video games.
11
Of course, there also are measurement reliability issues that will tend
to make empirical estimates smaller than true effects.
170
ANDERSON ET AL.
Magnitude of Average Effect Sizes
We have no doubt that the import of the average effect sizes
reported in this article will be greatly debated. There are at least
two major issues, one concerning which estimates are most appropriate and the other concerning the importance of the average
effect sizes.
Consider the results for physically aggressive behavior, shown
in Figure 1. Which are the most appropriate estimates of the true
violent video game effect, the raw estimates for cross-sectional and
longitudinal studies or those that partialed out sex (and Time 1)
effects? We believe that there is no single correct answer. On the
one hand, the two facts that males spend more time than females
playing violent video games and that males generally (but not
always) are more physically aggressive than females suggest that
an appropriately conservative test of violent video game effects
should partial out sex effects. On the other hand, once it has been
established that violent video game play is a causal risk factor for
physical aggression in both males and females, it becomes clear
that partialing out sex effects yields underestimates of the effect of
violent video games.
Concerning the importance issue, are the effect sizes large
enough to be considered important? From a theoretical standpoint,
the answer is pretty clear. Prevailing social– cognitive theories
predict statistically significant effects but do not predict absolute
magnitude; finding the predicted effects to be significant therefore
lends support to the theories.
From a practical or applied standpoint, the answer is less clear.
From a strict “percentage of variance explained” perspective, the
longitudinal effect size with sex and Time 1 aggression partialed
out might seem small (i.e., 0.1522 ⫽ 2.31%). However, as numerous authors have pointed out, even small effect sizes can be of
major practical significance. When effects accumulate across time,
or when large portions of the population are exposed to the risk
factor, or when consequences are severe, statistically small effects
become much more important (Abelson, 1985; Rosenthal, 1986,
1990). All three of these conditions apply to violent video game
effects.
Furthermore, when dealing with a multicausal phenomenon
such as aggression, one should not expect any single factor to
explain much of the variance. There are dozens of known risk
factors for both short-term aggression and the development of
aggression-prone individuals. To expect any one factor to account
for more than a small fraction of variance is unrealistic. This
suggests that a better way to assess the practical importance of
violent video game effects is to compare it to some other known
risk factors for youth violence. In fact, even the overly conservative sex and Time 1 adjusted VGV effect size estimate (r⫹ ⫽ .152)
compares favorably to such risk factors as substance use, abusive
parents, and poverty (U.S. Department of Health and Human
Services, 2001).12
Learning From the Past
Learning from our mistakes. Additional findings of importance come not from the meta-analytic results but from the review
and selection process involved in identifying studies of sufficient
quality to merit inclusion in the focal analyses. There are a lot of
methodological pitfalls that the wary researcher must avoid in this
domain, there are a lot of researchers who are not avoiding them,
and there are a lot of editors who are publishing them. The quality
criteria described in Table 2 and the more specific examples
outlined on our website could (and should) prove useful to many
people who are involved in media violence research, including
faculty supervisors of undergraduate and graduate students, journal
editors and their manuscript reviewers, and some media violence
scholars themselves.
We do not intend to imply that those studies that did not meet all
of our quality criteria are useless; some of them yielded findings of
some value on other research questions. Similarly, we do not
intend to imply that those studies that met our best practices
criteria are without shortcomings. There is no such thing as a
perfect study. Several of the most frequent weaknesses of studies
that did pass the present quality check are use of single-item
outcome measures; use of longitudinal time periods that may be
too short for the phenomenon of interest; use of weak measures of
exposure to violent video games; and sample sizes that are too
small given the expected effect size. The first three tend to lead to
underestimates of the true effect size, whereas the third tends to
yield nonsignificant statistical test results.
Contrary to claims by video game industry representatives,
some gamers, and a few researchers, in general it is not the
methodologically poor studies that tend to yield big effects.
Rather, methodologically superior studies tend to yield larger
effects.
Gaps and future research. This review also revealed a number of gaps in the research literature. One involves the lack of age
effects, especially in longitudinal studies. Most scholars in this
area believe that younger children are likely to be more vulnerable
than older adolescents and young adults to long-term media violence effects. But there is little evidence of larger effect sizes for
younger than for older participants. Indeed, a recent television
study suggests that even late adolescents are vulnerable (Johnson,
Cohen, Smailes, Kasen, & Brook, 2002).
Our cross-sectional analysis yielded a marginally significant
negative relationship between age and the effect of violent games
on aggressive behavior, but the longitudinal studies did not find an
age effect. Part of the problem may be that meta-analytic procedures are not ideally suited to testing such effects. Different studies
typically use different measures of aggression and video game
habits, include different participant populations, and take place
during different years. All of the differences add noise to the
meta-analysis, noise that may well hide real age effects. New
research is needed in which different-age participants are sampled
from the same population, the same measures are used (or as
similar as is reasonable given the age differences), and all of the
data collection takes place in the same years. Such studies should
yield cleaner tests of whether or not long-term effects of video
game violence are larger for younger than older participants.
12
It is unclear from this Surgeon General report whether all of the
reported longitudinal effects were similarly based on sex and Time 1
adjusted partial correlations. Also note that one criterion for inclusion in
the Surgeon General report was violent behavior. The present longitudinal
studies of the VGV/physical aggression include some violent behavior but
also include somewhat less severe forms of physical aggression. Thus, the
effects are not strictly comparable.
VIOLENT VIDEO GAME EFFECTS
Similarly, we believe that many moderator variables of interest
can be more precisely examined within the same study than they
can in meta-analyses. For example, as a test for short-term player
perspective effects, it would be best to conduct experiments in
which participants are randomly assigned to play the same violent
game but from different perspectives. Similar studies are needed to
more closely examine player role effects, target effects, and time
effects.
In the aggression domain, pre–post and repeated measures designs are generally problematic because of suspicion, practice, and
carryover effects. Longitudinal designs solve these problems to
some extent, but such designs do not allow experimental tests of
the immediate, short-term consequences of playing violent video
games. Some types of dependent variables are less susceptible to
these problems, such as reaction time tasks that presumably assess
effects without the participant’s awareness (e.g., lexical decision,
reading reaction time, and implicit association tasks). Cognitive
neuroscience approaches also are likely to prove especially valuable in more precisely assessing the immediate effects of playing
violent and nonviolent video games on a host of cognitive and
affective processes (e.g., Bailey, West, & Anderson, in press;
Weber, Ritterfeld, & Mathiak, 2006). For example, previous research has mapped specific brain areas that are especially active
during violent and nonviolent actions within the same game, using
functional magnetic resonance imaging techniques (Weber et al.,
2006).
Conclusions
The present findings show that the social– cognitive theoretical
view fits the existing data on video game violence effects quite
well. This has important implications for public policy debates, for
further development and testing of basic theory, and for development and testing of potential intervention strategies designed to
reduce harmful effects of playing violent video games. Concerning
basic theory, additional research of all three types (but especially
experimental and longitudinal) is needed, especially on VGV
effects on empathy, desensitization, and prosocial behavior. Additional longitudinal studies with longer intervals are needed for
aggressive behavior and aggressive cognition. Furthermore, longitudinal studies with very large samples and very long time spans
between the first time period and the last are needed so we can
assess the impact of violent video games on very serious forms of
physical aggression (i.e., violence). Concerning interventions,
there have been a few studies with findings that suggest that
specific programs involving schoolchildren and their parents can
reduce exposure to violent media and the frequency of unwarranted aggressive behavior (e.g., Huesmann, Eron, Klein, Brice, &
Fischer, 1983; Robinson, Wilde, Navracruz, Haydel, & Varady,
2001).
Concerning public policy, we believe that debates can and
should finally move beyond the simple question of whether violent
video game play is a causal risk factor for aggressive behavior; the
scientific literature has effectively and clearly shown the answer to
be “yes.” Instead, we believe the public policy debate should move
to questions concerning how best to deal with this risk factor.
Public education about this risk factor—and about how parents,
schools, and society at large can deal with it— could be very
useful.
171
It is true that as a player you are “not just moving your hand on
a joystick” but are indeed interacting “with the game psychologically and emotionally.” It is not surprising that when the game
involves rehearsing aggressive and violent thoughts and actions,
such deep game involvement results in antisocial effects on the
player. Of course, the same basic social– cognitive processes
should also yield prosocial effects when game content is primarily
prosocial. Unfortunately, there has been relatively little research
on purely prosocial game effects, largely because there are few
games that have the main characters modeling helpful behavior in
the complete absence of violent behavior. However, some recent
studies have found that prosocial games can increase cooperation
and helping (Gentile et al., 2009; Greitemeyer & Osswald, 2009).
Video games are neither inherently good nor inherently bad. But
people learn. And content matters.
References
Abelson, R. P. (1985). A variance explanation paradox: When a little is a
lot. Psychological Bulletin, 97, 129 –133.
Anderson, C. A. (2004). An update on the effects of violent video games.
Journal of Adolescence, 27, 113–122.
Anderson, C. A., Berkowitz, L., Donnerstein, E., Huesmann, L. R., Johnson, J., Linz, D., . . . Wartella, E. (2003). The influence of media
violence on youth. Psychological Science in the Public Interest, 4,
81–110.
Anderson, C. A., & Bushman, B. J. (2001). Effects of violent video games
on aggressive behavior, aggressive cognition, aggressive affect, physiological arousal, and prosocial behavior: A meta-analytic review of the
scientific literature. Psychological Science, 12, 353–359.
Anderson, C. A., & Bushman, B. J. (2002). Human aggression. Annual
Review of Psychology, 53, 27–51.
Anderson, C. A., & Carnagey, N. L. (2004). Violent evil and the general
aggression model. In A. Miller (Ed.), The social psychology of good and
evil (pp. 168 –192). New York, NY: Guilford.
Anderson, C. A., Carnagey, N. L., Flanagan, M., Benjamin, A. J., Eubanks,
J., & Valentine, J. C. (2004). Violent video games: Specific effects of
violent content on aggressive thoughts and behavior. Advances in Experimental Social Psychology, 36, 199 –249.
Anderson, C. A., & Dill, K. E. (2000). Video games and aggressive
thoughts, feelings, and behavior in the laboratory and in life. Journal of
Personality and Social Psychology, 78, 772–790.
Anderson, C. A., Gentile, D. A., & Buckley, K. E. (2007). Violent video
game effects on children and adolescents: Theory, research, and public
policy. New York, NY: Oxford University Press.
Anderson, C. A., & Huesmann, L. R. (2003). Human aggression: A
social– cognitive view. In M. A. Hogg & J. Cooper (Eds.), Handbook of
social psychology (pp. 296 –323). London, England: Sage.
Anderson, C. A., Lindsay, J. J., & Bushman, B. J. (1999). Research in the
psychological laboratory: Truth or triviality? Current Directions in
Psychological Science, 8, 3–9.
Bachrach, R. (1986). The differential effect of observation of violence on
kibbutz and city children in Israel. In L. R. Huesmann & L. D. Eron
(Eds.), Television and the aggressive child: A cross-national comparison
(pp. 210 –238). Hillsdale, NJ: Erlbaum.
Bailey, K., West, R., & Anderson, C. A. (in press). The influence of video
games on social, cognitive, and affective information processing. In J.
Decety & J. Cacioppo (Eds.), Handbook of social neuroscience. New
York, NY: Oxford University Press.
Bandura, A. (1973). Aggression: A social learning theory analysis. Englewood Cliffs, NJ: Prentice-Hall.
Bandura, A., Barbaranelli, C., Caprara, G. V., & Pastorelli, C. (1996).
172
ANDERSON ET AL.
Mechanisms of moral disengagement in the exercise of moral agency.
Journal of Personality and Social Psychology, 71, 364 –374.
Barclay, G., & Tavares, C. (2002). International comparisons of criminal
justice statistics 2000. Retrieved from United Kingdom Home Office
website: http://www.homeoffice.gov.uk/rds/pdfs2/hosb502.pdf
Bartholow, B. D., Bushman, B. J., & Sestir, M. A. (2006). Chronic violent
video game exposure and desensitization to violence: Behavioral and
event-related brain potential data. Journal of Experimental Social Psychology, 42, 532–539.
Bergeron, N., & Schneider, B. H. (2005). Explaining cross-national differences in peer-directed aggression: A quantitative synthesis. Aggressive Behavior, 31, 116 –137.
Berkowitz, L. (1984). Some effects of thoughts on anti- and prosocial
influences of media events: A cognitive–neoassociation analysis. Psychological Bulletin, 95, 410 – 427.
Berkowitz, L. (1993). Aggression: Its causes, consequences, and control.
New York, NY: McGraw-Hill.
Bushman, B. J., & Anderson, C. A. (1998). Methodology in the study of
aggression: Integrating experimental and nonexperimental findings. In
R. Geen & E. Donnerstein (Eds.), Human aggression: Theories, research, and implications for social policy (pp. 23– 48). San Diego, CA:
Academic Press.
Bushman, B. J., & Anderson, C. A. (2009). Comfortably numb: Desensitizing effects of violent media on helping others. Psychological Science,
20, 273–277.
Bushman, B. J., & Huesmann, L. R. (2006). Short-term and long-term
effects of violent media on aggression in children and adults. Archives of
Pediatric and Adolescent Medicine, 160, 348 –352.
Buss, A. H., & Perry, M. (1992). The Aggression Questionnaire. Journal
of Personality and Social Psychology, 63, 452– 459.
Carlson, M., Marcus-Newhall, A., & Miller, N. (1989). Evidence for a
general construct of aggression. Personality and Social Psychology
Bulletin, 15, 377–389.
Carnagey, N. L., Anderson, C. A., & Bushman, B. J. (2007). The effect of
video game violence on physiological desensitization to real-life violence. Journal of Experimental Social Psychology, 43, 489 – 496.
Chida, Y., & Hamer, M. (2008). Chronic psychosocial factors and acute
physiological responses to laboratory-induced stress in healthy populations: A quantitative review of 30 years of investigations. Psychological
Bulletin, 134, 829 – 885.
Colwell, J., & Kato, M. (2003). Investigation of the relationship between
social isolation, self-esteem, aggression and computer game play in
Japanese adolescents. Asian Journal of Social Psychology, 6, 149 –158.
Comstock, G., & Scharrer, E. (2007). Media and the American child. San
Diego, CA: Academic Press.
Crick, N. R., & Dodge, K. A. (1994). A review and reformulation of social
information processing mechanisms in children’s adjustment. Psychological Bulletin, 115, 74 –101.
DeLisi, M. (2005). Career criminals in society. Thousand Oaks, CA: Sage.
Duval, S. (2005). The “Trim and Fill” method. In H. Rothstein, A. Sutton,
& M. Borenstein (Eds.), Publication bias in meta-analysis: Prevention,
assessment and adjustments (pp. 127–144). Chichester, England: Wiley.
Duval, S. J., & Tweedie, R. L. (1998). Practical estimates of the effect of
publication bias in meta-analysis. Australasian Epidemiologist, 5(4),
14 –17.
Duval, S., & Tweedie, R. (2000a). A nonparametric “Trim and Fill”
method of accounting for publication bias in meta-analysis. Journal of
the American Statistical Association, 95, 89 –98.
Duval, S., & Tweedie, R. (2000b). Trim and fill: A simple funnel-plotbased method of testing and adjusting for publication bias in metaanalysis. Biometrics, 56, 455– 463.
Ferguson, C. J. (2007a). Evidence for publication bias in video game
violence effects literature: A meta-analytic review. Aggression and
Violent Behavior, 12, 470 – 482.
Ferguson, C. J. (2007b). The good, the bad and the ugly: A meta-analytic
review of positive and negative effects of violent video games. Psychiatric Quarterly, 78, 309 –316.
Ferguson, C. J., & Kilburn, J. (2009). The public health risks of media
violence: A meta-analytic review. Journal of Pediatrics, 154, 759 –763.
Frijda, N. H., Markam, S., Sato, K., & Wiers, R. (1995). Emotion and
emotion words. In J. A. Russell (Ed.), Everyday conceptions of emotion
(pp. 121–143). Dordrecht, the Netherlands: Kluwer Academic.
Geen, R. G. (2001). Human aggression. Philadelphia, PA: Open University
Press.
Gentile, D. A. (Ed.). (2003). Media violence and children. Westport, CT:
Praeger.
Gentile, D. A., Anderson, C. A., Yukawa, S., Ihori, N., Saleem, M., Ming,
L. K., . . . Sakamoto, A. (2009). The effects of prosocial video games on
prosocial behaviors: International evidence from correlational, experimental, and longitudinal studies. Personality and Social Psychology
Bulletin, 35, 752–763.
Gentile, D. A., Saleem, M., & Anderson, C. A. (2007). Public policy and
the effects of media violence on children. Social Issues and Policy
Review, 1, 15– 61.
Giancola, P. R., & Chermack, S. T. (1998). Construct validity of laboratory
aggression paradigms: A response to Tedeschi and Quigley (1996).
Aggression and Violent Behavior, 3, 237–253.
Giancola, P. R., & Parrott, D. J. (2008). Further evidence for the validity
of the Taylor aggression paradigm. Aggressive Behavior, 34, 214 –229.
Greitemeyer, T., & Osswald, S. (2009). Prosocial video games reduce
aggressive cognitions. Journal of Experimental Social Psychology, 45,
896 –900.
Huesmann, L. R. (1988). An information processing model for the development of aggression. Aggressive Behavior, 14, 13–24.
Huesmann, L. R. (1998). The role of social information processing and
cognitive schema in the acquisition and maintenance of habitual aggressive behavior. In R. G. Geen & E. Donnerstein (Eds.), Human aggression: Theories, research, and implications for social policy (pp. 73–
109). New York, NY: Academic Press.
Huesmann, L. R., & Eron, L. D. (Eds.). (1986). Television and the aggressive child: A cross-national comparison. Hillsdale, NJ: Erlbaum.
Huesmann, L. R., Eron, L. D., Klein, R., Brice, P., & Fischer, P. (1983).
Mitigating the imitation of aggressive behaviors by changing children’s
attitudes about media violence. Journal of Personality and Social Psychology, 44, 899 –910.
Huesmann, L. R., Lagerspetz, K., & Eron, L. D. (1984). Intervening
variables in the TV violence–aggression relation: Evidence from two
countries. Developmental Psychology, 20, 746 –775.
Johnson, J. G., Cohen, P., Smailes, E. M., Kasen, S., & Brook, J. S. (2002,
July 5). Television viewing and aggressive behavior during adolescence
and adulthood. Science, 295, 2468 –2471.
Kirsh, S. J. (2006). Children, adolescents, and media violence: A critical
look at the research. Thousand Oaks, CA: Sage.
Kitayama, S., Mesquita, B., & Karasawa, M. (2006). Cultural affordances
and emotional experience: Socially engaging and disengaging emotions
in Japan and the United States. Journal of Personality and Social
Psychology, 91, 890 –903.
Kodaira, S. I. (1998). A review of research on media violence in Japan. In
U. Carlsson & C. von Feilitzen (Eds.), Children and media violence (pp.
81–105). Göteborg, Sweden: UNESCO International Clearinghouse on
Children and Violence on the Screen.
Kodomo no taiken katsudo kenkyukai. (2000). Kodomo no taiken katsudo
tou ni kansuru kokusai hikaku chosa [A cross-cultural study about
children’s experiences and activities]. Nagano, Japan: Author.
Krahé, B. (2001). The social psychology of aggression. East Sussex,
England: Psychology Press.
Lane, K. A., Banaji, M. R., Nosek, B. A., & Greenwald, A. G. (2007).
Understanding and using the Implicit Association Test: IV. What we
VIOLENT VIDEO GAME EFFECTS
know (so far) about the method. In B. Wittenbrink & N. Schwarz (Eds.),
Implicit measures of attitudes (pp. 59 –102). New York, NY: Guilford
Press.
Masuda, T., & Nisbett, R. E. (2001). Attending holistically versus analytically: Comparing the context sensitivity of Japanese and Americans.
Journal of Personality and Social Psychology, 81, 922–934.
Mesquita, B., & Leu, J. (2007). The cultural psychology of emotion. In S.
Kitayama & D. Cohen (Eds.), Handbook of cultural psychology (pp.
734 –759). New York, NY: Guilford Press.
Nisbett, R. E., Peng, K., Choi, I., & Norenzayan, A. (2001). Culture and
systems of thought: Holistic versus analytic cognition. Psychological
Review, 108, 291–310.
Ontario Office for Victims of Crime. (2004). Action agenda: A strategic
blueprint for reducing exposure to media violence in Canada. Retrieved
from http://www.fradical.ca/Action_Agenda_November_2004.pdf
Polman, H., Orobio de Castro, B., & Van Aken, M. A. G. (2008). Experimental study of the differential effects of playing versus watching
violent video games on children’s aggressive behavior. Aggressive Behavior, 34, 256 –264.
Robinson, T. N., Wilde, M. L., Navracruz, L. C., Haydel, K. F., & Varady,
A. (2001). Effects of reducing children’s television and video game use
on aggressive behavior: A randomized controlled trial. Archives of
Pediatric Adolescent Medicine, 155, 17–23.
Rosenthal, R. (1986). Media violence, antisocial behavior, and the social
consequences of small effects. Journal of Social Issues, 42, 141–154.
Rosenthal, R. (1990). How are we doing in soft psychology? American
Psychologist, 45, 775–777.
Rothstein, H. R., Sutton, A. J., & Borenstein, M. (2005). Publication bias
in meta-analysis: Prevention, assessment and adjustments: Chichester,
England: Wiley.
Schmitt, D. P., Allik, J., McCrae, R. R., & Benet-Martinez, V. (2007). The
geographic distribution of Big Five personality traits: Patterns and
profiles of human self-description across 56 nations. Journal of CrossCultural Psychology, 38, 173–212.
Sherry, J. L. (2001). The effects of violent video games on aggression: A
meta-analysis. Human Communication Research, 27, 409 – 431.
173
Singer, D. G., & Singer, J. L. (Eds.). (2001). Handbook of children and the
media. Thousand Oaks, CA: Sage.
Sterne, J. A. C., & Egger, M. (2005). Regression methods to detect
publication and other bias in meta-analysis. In H. Rothstein, A. Sutton,
& M. Borenstein (Eds.), Publication bias in meta-analysis: Prevention,
assessment and adjustments (pp. 99 –110). Chichester, England: Wiley.
Strack, F., & Deutsch, R. (2004). Reflective and impulsive determinants of
social behavior. Personality and Social Psychology Review, 8, 220 –247.
Taylor, S. P. (1967). Aggressive behavior and physiological arousal as a
function of provocation and the tendency to inhibit aggression. Journal
of Personality, 35, 297–310.
Tsai, J. L., Knutson, B., & Fung, H. H. (2006). Cultural variation in affect
valuation. Journal of Personality and Social Psychology, 90, 288 –307.
U.S. Department of Health and Human Services. (2001). Youth violence: A
report of the Surgeon General. Rockville, MD: U.S. Department of
Health and Human Services, Centers for Disease Control and Prevention, National Center for Injury Prevention and Control; Substance
Abuse and Mental Health Services Administration, Center for Mental
Health Services; and National Institutes of Health, National Institute of
Mental Health.
Weber, R., Ritterfeld, U., & Mathiak, K. (2006). Does playing violent
video games induce aggression? Empirical evidence of a functional
magnetic resonance imaging study. Media Psychology, 8, 39 – 60.
Wegner, D. M., & Bargh, J. A. (1998). Control and automaticity in social
life. In D. Gilbert, S. Fiske, & G. Lindzey (Eds.), The handbook of social
psychology (pp. 446 – 496). New York, NY: McGraw-Hill.
Williams, D., & Skoric, M. (2005). Internet fantasy violence: A test of
aggression in an online game. Communication Monographs, 72, 217–
233.
Yahiro, S. (2005). Terebigemu kaishakuron josetsu: Assanburaaju [Introduction of interpreting theory of video games: Assemblage]. Tokyo,
Japan: Gendaishokan.
Received July 6, 2009
Revision received October 17, 2009
Accepted October 21, 2009 䡲